Engineering Applications of Artificial Intelligence 129 (2024) 107597

Contents lists available at ScienceDirect

Engineering Applications of Artificial Intelligence

journal homepage: www.elsevier.com/locate/engappai

MEDS-Net: Multi-encoder based self-distilled network with bidirectional

maximum intensity projections fusion for lung nodule detection
Muhammad Usman a,b,c ,∗, Azka Rehman b,d , Abdullah Shahid b , Siddique Latif e , Yeong-Gil Shin a
a
Department of Computer Science and Engineering, Seoul National University, 1 Gwanak-ro, Gwanak-gu, Seoul, 08826, South Korea
b
Center for Artificial Intelligence in Medicine and Imaging, HealthHub Co. Ltd., Seoul, 06524, South Korea
c
SeoulDynamics Inc., Seongnam-si, Gyeonggi-do, 13449, South Korea
d
Department of Biomedical Sciences, Seoul National University Graduate School, Seoul, 03080, South Korea
e
School of Computer Science and Engineering, University of New South Wales, Sydney NSW 2052, Australia

ARTICLE INFO ABSTRACT

Keywords: Early detection of lung cancer increases cure rates, making precise detection of lung nodules in computed
Bidirectional maximum intensity projection tomography (CT) images crucial. However, creating an accurate computer-aided detection (CADe) system is
Computer-aided detection (CADe) challenging due to lung structure complexity and nodule heterogeneity. We present a lung nodule detection
Multi-encoder network
scheme, inspired by radiologist workflow, fusing CT scan sub-volumes with bidirectional maximum intensity
Pulmonary nodule detection
projection (MIP) images within a single architecture. We have developed a novel multi-encoder-based network
Self-distillation
(MEDS-Net) using self-distillation to learn from three types of inputs: 3D sub-volume, forward, and backward
MIP images. MEDS-Net utilizes three encoders to extract insights, passed to the decoder via attention units to
suppress redundant features. The decoder employs a self-distillation mechanism, connecting to a distillation
block with four auxiliary lung nodule detectors, accelerating convergence and enhancing learning ability. The
auxiliary detectors are used during inference to reduce false positives. Our scheme, tested on 888 LIDC-IDRI
dataset scans, achieved a 93.6% competition performance metric score, illustrating that fusing bidirectional
MIP images with raw CT slices and auxiliary detectors empowers MEDS-Net to accurately detect lung nodules
while minimizing false positives.

1. Introduction employed in 2D and 3D to achieve significantly improved perfor-

Lung cancer is the most frequent cancer type, with the highest
mortality rates globally (Sung et al., 2021). The survival rate of patients
highly depends upon the early detection of pulmonary nodules (Bald-
win, 2015). Computed tomography (CT) is an extensively employed
effective medical imaging modality for diagnosing pulmonary nod-
ules (Sluimer et al., 2006). However, it is often a time-consuming
and tedious job to manually identify nodules in CT scans. Radiologists
need to read the CT scans slice by slice, and depending upon the slice
thickness, a chest CT may contain hundreds of slices. Additionally,
manual lung nodule detection is error-prone (Seemann et al., 1999).
To address this problem, computer-aided detection (CADe) systems
becoming popular in aiding radiologists in accurately diagnosing lung
cancer.
Numerous efforts (e.g., Zhao et al. (2022), Zhou et al. (2022))
have been made in designing CADe systems. Recently, deep learning
(DL) based techniques have made vast inroads in lung nodule de-
tection (Zhang et al., 2018; Halder et al., 2020). In DL-based CADe
systems, convolutional neural networks (CNNs) have been extensively
mance (Monkam et al., 2019). Such CADe system typically consists of
two stages: screening and false positives (FPs) reduction. The former is
typically applied to identify suspicious regions (i.e., nodule candidates)
in a patient’s exam. The latter stage is to reduce the FPs by applying
the additional CNNs (Cao et al., 2020; Pezeshk et al., 2018; Zhao et al.,
2022; Zhou et al., 2022). Although such CADe systems demonstrated
promising performance for lung nodule detection, these techniques add
computational complexity by increasing the number of CNNs to train
and are more prone to error as nodules missed at the first stage cannot
be detected at the second stage. Subsequently, the sensitivity at the
detection stage is kept high, significantly increasing the FPs which adds
the load to the FPs reduction stage. To overcome the challenges as-
sociated with dual-stage-based CADe systems, single-stage frameworks
were proposed to detect the lung nodules (Li and Fan, 2020; Zhu
et al., 2022; Luo et al., 2022). These frameworks considerably improve
the computational complexities and ease the training and optimization
process; however, they suffer from a low performance that limits their
real-time clinical employment.

∗ Corresponding author at: Department of Computer Science and Engineering, Seoul National University, 1 Gwanak-ro, Gwanak-gu, Seoul, 08826, South Korea.
E-mail address: ussman@snu.ac.kr (M. Usman).

https://doi.org/10.1016/j.engappai.2023.107597
Received 19 July 2023; Received in revised form 23 September 2023; Accepted 20 November 2023
Available online 4 December 2023
0952-1976/© 2023 Elsevier Ltd. All rights reserved.
M. Usman et al.
Despite the advancements in performance, a significant portion of
existing CADe systems do not align with the clinical workflow, leading
to challenges in lung nodule detection. Without incorporating domain-
specific knowledge, these systems often resort to more intricate models
applied across multiple stages for accurate nodule identification. In
typical clinical practice, radiologists predominantly use the maximum
intensity projection (MIP) images (Wallis et al., 1989) of CT scans
for initial nodule screening, turning to the raw CT slices for final
confirmation. Notably, only a handful of studies (Masood et al., 2020;
Zheng et al., 2019) have integrated MIP images to enhance lung nod-
ule detection, aligning more closely with clinical practices. Although
these studies have demonstrated the value of aligning with clinical
workflows using relatively simpler, conventional networks, they have
often employed multiple networks across different stages. This multi-
network approach not only adds complexity but also presents hurdles
for real-time clinical deployment. To this end, our work introduces
a cohesive, single-stage CADe system that synergizes the 3D patch
of CT scans with MIP images for precise lung nodule detection. In
a pioneering effort to minimize training overhead while enhancing
performance, our methodology incorporates self-distillation into our
proposed architecture for lung nodule detection.
Below is a summary of the main contributions of this study in
contrast to the existing literature.
1.1. Summary
1. We design a novel single-stage multi-encoder-based CNN that
utilizes the self-distillation mechanism and improves the perfor-
mance compared to state-of-the-art single-stage as well as most
dual-stage frameworks on the public LIDC-IDRI dataset.
2. We propose to utilize bidirectional MIP images of three slab
thicknesses, i.e., 3, 5, and 10 mm, for lung nodule detection,
which improves the network’s ability to distinguish nodules from
non-nodular structures in the lung region.
3. To reduce false positives, we introduce our proposed method,
which uniquely leverages outputs from auxiliary branches of the
same architecture for both lung nodule candidate detection and
subsequent false positive reduction. This approach stands out
as it effectively utilizes the same network in a dual capacity, a
feature not commonly observed in existing methods.
2. Related work
2.1. Dual stage methods for lung nodule detection
Most DL-based computer-aided detection (CADe) systems for lung
nodule detection consist of two stages: screening and false positive
(FP) reduction. The former is typically applied to identify suspicious
regions (i.e., nodule candidates) in a patient’s exam. The latter stage
is to reduce the FPs by using the additional network (e.g., CNNs). For
instance, Cao et al. (2020) proposed a two-stage CADe system with an
improved UNet architecture for candidate deduction.
They used a two-phase inference scheme, i.e., rough segmentation
and fine segmentation in the smaller local region. To eliminate the
false positives, they use 3D-CNNs-based ensemble architecture. Pezeshk
et al. (2018) applied 3D-CNNs for the initial screening of lung nod-
ules, and FPs are reduced by an ensemble of 3D-CNNs trained using
different transformations applied to both the positive and negative
patches to augment the training set. Zhao et al. (2022) employed the
high-resolution fused attention module, containing channel and spatial
attention, to incorporate the global and spatial information of focal
nodules for detecting the nodule candidates. In the second stage, an
adaptive 3D CNN structure is designed to further reduce the false
positives, which extracts the multilevel contextual information via an
adaptive 3D convolution kernel.
2
Engineering Applications of Artificial Intelligence 129 (2024) 107597
Yuan et al. (2021) presented a dual-stage scheme that utilized 3-D
residual UNet with channel attention and applied multi-task learning
with a multi-branch classifier network for nodule detection and FPs
reduction, respectively. Pereira et al. (2021) used mask region-mask
region-convolutional neural network (Mask R-CNN) to detect bounding
boxes in axial slices and used a classifier ensemble based on CT attenua-
tion patterns to boost 3D pulmonary nodule classification performance.
Mei et al. (2021) used a 3D region proposal network to generate
pulmonary nodule candidates, and the multi-scale feature maps were
used to reduce the FPs. These dual-stage methods are computationally
expensive to run, and misdiagnosis and omission errors in the preceding
stages significantly impact the FP reduction in the subsequent stage. In
contrast, we propose a single-stage network for lung nodule detection.
2.2. Single stage methods for lung nodule detection
Due to the issues associated with dual-stage CADe systems, some
studies proposed single-stage CADe systems that simultaneously per-
form nodule detection and false positive reduction using a single net-
work. For instance, Li and Fan (2020) used the 3D encoder–decoder
architecture for lung nodule detection. To effectively reduce the false
positives, they utilize a dynamically scaled cross-entropy loss function.
Wu et al. (2021) combined image enhancement and a dual-branch neu-
ral network for improving the visibility of lung nodules by suppressing
the noises from the background. The enhanced images were fed to the
architecture based on two UNets to detect the nodules. Zhu et al. (2022)
utilized 3D residual UNet with improved attention gates to reduce the
false positives. They also applied a channel interaction unit prior to the
detection head and the gradient harmonizing mechanism loss function
to combat the problem of data imbalance. Luo et al. (2022) detected
the position, radius, and offset of nodules by using an anchor-free 3D
sphere representation-based center-points matching detection network.
Single-stage techniques greatly reduce the computational complexities
and are easier to train and optimize; however, they suffer from lower
performance, limiting their clinical usage.
2.3. MIP for lung nodule detection
In the real-time clinical workflow, mostly, radiologists first examine
the maximum intensity projection (MIP) images (Wallis et al., 1989) of
CT scan to roughly screen the nodule candidates and use the raw CT
scan slices to finalize the decision. MIP helps to summarize the shape
variations appearing along the z-axis, which improves the visibility
of lung nodules by distinguishing them from surrounding structures
inside the lung. It can be taken in both directions (i.e., forward and
backward), and each highlights the structural variations occurring in
each direction. Very few studies utilized the MIP images to improve
lung nodule detection performance and comply with the clinical work-
flow. For instance, Masood et al. (2020) leveraged MIP to extract
meaningful insights of lung nodules from 2D views (i.e., axial, coronal,
and sagittal) and used multidimensional region-based fully convolu-
tional neural network based CAD system for lung nodule detection and
classification, respectively. Similarly, Zheng et al. (2019) attempted to
follow the clinical workflow by MIP with different slab thicknesses
(i.e., 5 mm, 10 mm, 15 mm) to train four different UNet architectures
and merged their outputs for nodule candidates detection. They trained
two different-sized DNNs, which classified the 3D candidate patch into
nodule or non-nodule for false positive reduction. Based on the results,
they show that the performance of CADe for nodule detection can be
improved by leveraging MIP images. However, this study utilized four
MIP images with four different networks, each network was provided
with partial information, which is contrary to the conventional clinical
workflow. In clinical practices, the radiologist leverages raw CT slices
along with MIP images of various thicknesses to effectively detect lung
nodules (Guleryuz Kizil et al., 2020). It is infeasible for radiologists to
solely rely on MIP images for diagnosis as it provides the highlights of
M. Usman et al. Engineering Applications of Artificial Intelligence 129 (2024) 107597

Table 1
Summary of comparison of our work with the existing studies.
Author (Year) Input type False positive Single stage Self-
reduction or network distillation/auxiliary
outputs
2D/3D slices MIP images
Forward Backward
Pezeshk et al. (2018) ✓ ✗ ✗ ✓ ✗ ✗
Zheng et al. (2019) ✓ ✓ ✗ ✓ ✗ ✗
Cao et al. (2020) ✓ ✗ ✗ ✓ ✗ ✗
Li and Fan (2020) ✓ ✗ ✗ ✗ ✓ ✗
Masood et al. (2020) ✓ ✓ ✗ ✗ ✓ ✗
Pereira et al. (2021) ✓ ✗ ✗ ✓ ✗ ✗
Zhou et al. (2022) ✓ ✗ ✗ ✓ ✗ ✗
Zhu et al. (2022) ✓ ✗ ✗ ✗ ✓ ✗
Mei et al. (2021) ✓ ✗ ✗ ✓ ✗ ✗
Luo et al. (2022) ✓ ✗ ✗ ✗ ✓ ✗
Zhao et al. (2022) ✓ ✗ ✗ ✓ ✗ ✗
Ours (2023) ✓ ✓ ✓ ✓ ✓ ✓

adjacent slices by suppressing the lower intensities, and in some cases,
it also subdue the nodules lying at the lower intensity values. Therefore,
radiologists examine the raw CT scans complementary with MIP images
of various thicknesses to draw accurate decisions on lung nodules (Gru-
den et al., 2002). More particularly, it has proven to be more effective
for small nodules and nodules with higher malignancy (Jabeen et al.,
2019). Nevertheless, in this work, we completely emulate the clinical
workflow by simultaneously incorporating the raw CT slices and MIP
images of various thicknesses into proposed multi-encoder based self-
distilled network (MEDS-Net). We further introduce the novel concept
of bidirectional MIP images which enables MEDS-Net to exploit the
enriched insights to detect the lung nodules.
1. While some studies utilize forward MIP for lung nodule detec-
tion, however, we are the first to exploit bidirectional (forward
and backward) MIP images to boost the detection performance.
2. None of the studies used self-distillation in CADe systems for
lung nodule detection. Our model equips with self-distillation to
better optimize the model.
3. None of the previous work on single-stage models achieved bet-
ter performance compared to the dual-stage techniques. Whereas
or model utilizes auxiliary outputs from various levels of the de-
coder block to reduce false positives in a single stage and achieve
state-of-the-art performance compared to dual-stage models.

2.4. Self-distillation for lung nodule detection 3. Proposed methodology

The conventional knowledge distillation (KD) used a pre-trained
teacher model to produce the soft labels as ground truth to train
a student model (Hinton et al., 2015). It was originally introduced
to compress a big network (teacher) to a small network (student),
i.e., network compression (Hui et al., 2018; Tung and Mori, 2019).
It computes a distillation loss between teacher and student, for which
the implementation methods can be logits-based (Jiang et al., 2019;
Phuong and Lampert, 2019), feature-based (Liu et al., 2020; Xu et al.,
2020b), and hybrid (Li et al., 2017; Xu et al., 2020a). For instance, Qin
et al. (2021) feature-based KD for the segmentation task in the medical
images. In recent works, KD has been extended to its self-learning
version called self-distillation in which teacher and student networks
share the same backbone architecture. Self-distillation helps models
converge to flat minima which features in generalization inherently,
prevents models from vanishing gradient problem, and enables the
network to learn more discriminating features (Zhang et al., 2019).
Subsequently, it has achieved impressive results in several visual tasks,
such as image classification (Zhang et al., 2019; Luan et al., 2019;
Xu et al., 2020b) and image segmentation (Wang et al., 2022; Ni
et al., 2022). Most importantly, self-distillation has also been proven
effective for improving the performance of the main (teacher) model
for medical image segmentation (Dong et al., 2022; Xie et al., 2022).
These schemes perform layer-wise context distillation to optimize the
shallow networks during the training process which eventually helps
the deepest architecture to obtain better performance. However, these
techniques do not utilize shallow networks during the inference process
despite having same backbone as the deepest network. In contrast to
the previous techniques in this work, we not only employ the self-
distillation during the training process to improve the lung nodule
detection ability of the proposed MEDS-Net but also at the inference,
we exploit the shallow networks, called auxiliary detectors, to reduce
the false positives.
We highlight the difference between our work with the existing
work in Table 1.
In this section, we describe the proposed framework for lung nodule
detection in abdominal CT scans, which mimics the clinical prac-
tices of radiologists who utilize maximum intensity projection (MIP)
images (Wallis et al., 1989). Our proposed multi-encoder-based self-
distilled network (MEDS-Net) is designed to fully exploit the MIP
images along with a 3D patch of CT scan as shown in Fig. 1.
The proposed MEDS-Net leverages the self-distillation mechanism
and auxiliary outputs to optimize the learning process and reduce
false positives at the inference stage, respectively. The details of each
component of the proposed framework have been described in the
following subsections.
3.1. Bidirectional maximum intensity projection
Maximum intensity projection (MIP) image consists of maximum
values, which come across the plane of projection, i.e., the 𝑧-axis. MIP
images of thoracic CT scans have proven effective for lung nodule
detection (Gruden et al., 2002) and it has been extensively utilized
by radiologists to quickly screen lung nodules. MIP images can be
generated of various slab thicknesses; however, research has shown that
slab thicknesses up to 10 mm are the most effective ones, and increasing
the slab thickness further reduces the detection sensitivity (Zheng et al.,
2020). Subsequently, in this work, we utilize MIP images with three
slab thicknesses, i.e., 3 mm, 5 mm, and 10 mm. The generation of a
MIP image can be described as follows:
𝐼𝑥,𝑦 = max 𝐷𝑥,𝑦,𝛥𝑧 (1)
𝛥𝑧
where, 𝐼𝑥,𝑦 and 𝐷𝑥,𝑦,𝛥𝑧 represent the generated MIP image and sub-
volume or slab of CT scan, respectively. Here, 𝛥𝑧 denotes the slab
thickness, which in our case is set to 3, 5, and 10 mm. Traditionally,
MIP is taken in one direction, which helps to distinguish nodules from
the vessels penetrating in one direction only. In this work, we take
it one step further to improve the visibility of nodules by utilizing

3
M. Usman et al.
Fig. 1. The illustration of our proposed computer-aided detection system. Multi-encoder based self-distilled network (MEDS-Net) which incorporates three different inputs, i.e., 3D
sub-volume of CT scan, forward and backward MIP images of three slab thickness (i.e., 3, 5 and 10 mm).
the bidirectional MIP. Therefore, the range of 𝛥𝑧 is defined differently
while generating the forward and backward MIP image which can be
formulated as follows:
{[ ] Net) for end-to-end lung nodule detection framework is demonstrated
𝜙𝑐 , 𝜙𝑐 + 𝜙𝑚 , For forward MIP image
𝛥𝑧 = [ ] (2)
𝜙𝑐 − 𝜙𝑚 , 𝜙𝑐 , For backward MIP image
where, 𝜙𝑐 and 𝜙𝑚 represent the depth of the current slice in the scan
and the thickness of MIP which has to be generated, i.e., 3, 5, and
10 mm, respectively. The process of generating the bidirectional MIP
images, in forward and backward directions, with three slab thick-
nesses, has been visually demonstrated in Fig. 2(a). Fig. 2(b) shows the
changes in the appearance of nodules in forward and backward MIP
images. Increasing the thickness of MIP image improves the visibility
of vessels which helps to distinguish the nodule from other non-nodular
structures. The forward and backward MIP of the same depth provide
different aspects of the nodule. To further elaborate the intuition of
introducing bidirectional MIP images, in Fig. 3, backward and forward
MIP images of 3 mm thickness with corresponding normal slices have
been shown for four selected nodules, i.e., (a), (b), (c), and (d). In all
the samples, the nodules’ appearance in the backward MIP is more
distinguishable, whereas, in the forward MIP, nodules appear more
connected with vessels which makes the detection harder. Therefore,
incorporating backward MIP images into the proposed MEDS-Net can
greatly complement the network by enhancing the nodule’s visibility.
Moreover, we also normalize the slice thickness along the 𝑧-axis of
the raw CT scan to 1 mm while keeping the original pixel spacing in
xy-plane for 3D patch-based input.
4
Engineering Applications of Artificial Intelligence 129 (2024) 107597
3.2. Multi-encoder based self-distilled network
Our proposed multi-encoder-based self-distilled network (MEDS-
in Fig. 1. We summarize its working as follows: (1) first 3D sub-volume
(stack of slices) consisting of a total of 11 slices (i.e., 11 × 256 × 256)
of 1 mm thickness is fed to a dense block which squeezes the repre-
sentation into three channels by using dense units; (2) the compressed
features from dense block along with forward and backward MIP
images are inputted into encoder block which contains three deep
encoders to extract the meaningful features from each three-channel
input; (3) the extracted features at various levels of three encoders
are concatenated to combine with the corresponding decoder layers
via attention units to decoder block; (4) decoder block utilizes the
deconvolutional layers to up-sample the latent features at each stage
that are utilized as gating signals to attention units. At five levels of the
decoder, the up-scaled refined features are inputted to the distillation
block and main detector to produce the detection results; (5) finally, the
main detector produces all the nodule candidates while the distillation
block utilizes four auxiliary detectors to reshape the refined multi-scale
features to the dimensions same as the input image, i.e., 256 × 256 for
acquiring the nodule candidates.
The details of each block of the proposed MEDS-Net have been
described in the following subsections.
3.2.1. Dense block
To incorporate the slice-level spatial and sequential information
that is crucial to detect the small nodules, we input a stack of slices,
M. Usman et al.
Fig. 2. Demonstration of bidirectional MIP images. (a) Process of generating the forward and backward MIP images. (b) Examples of forward and backward MIP of thicknesses
3 mm, 5 mm, and 10 mm (left to right).
called sub-volume, to the proposed architecture. Concretely, the five
adjacent slices from both sides of the central slice, i.e., the 3D sub-
volume with dimensions 11 × 256 × 256, is extracted and inputted
into a 3D dense block, which is shown in Fig. 4(a). 3D dense block
takes 11 slices of 1 mm thickness as input and transforms into the
meaningful latent representative of the same dimension as three MIP
images, i.e., 3 × 256 × 256. The 3D dense block is composed of four sets
of dense units followed by the max-pooling layers, and the fifth dense
unit is followed by the reshaping layer, which squeezes the information
to desired dimensions. Fig. 4(b) shows the architecture of each dense
unit which is composed of five sets of 3D convolution, Relu activation,
and batch normalization. The output of each set is propagated to all
the next sets using skip connections. After each dense unit features are
down-sampled along the depth, subsequently, at the end of the dense
block, we have three channels of feature maps. These feature maps are
further fed into the encoder block to extract the underlying insights
pertaining to nodule detection.
3.2.2. Encoder block
In Fig. 1, the overall structure of the encoder block has been
demonstrated, which includes three coding paths, namely, 3D sub-
volume, forward, and backward MIPs encoders. All three encoders are
kept identical and inspired by conventional UNet encoder (Ronneberger
et al., 2015). Nevertheless, our encoders are deeper than the encoder
in the standard UNet to effectively learn meaningful insights from
lung structure which are crucial to distinguish the lung nodules. Each
encoder consists of six Conv-Blocks which down-sample the input to
lower dimensions while extracting the meaningful features. As shown
in Fig. 5(a), Conv-Block has two stages; the first stage has two 2D
convolutional layers followed by a batch normalization layer, then a
skip connection is used to combine the input with the output features
of the first convolution stage. This type of short skip connection tends
to stabilize gradient updates (He et al., 2016). In the second stage,
the combined features are passed to the same set of layers as the
first stage, followed by an additional down-sampling layer. The width
5
Engineering Applications of Artificial Intelligence 129 (2024) 107597
of each of our branches is kept small to avoid overfitting. Our first
Conv-Block contains eight feature maps that get doubled after every
Conv-Block, and eventually, our final Conv-Block has 256 features.
Finally, the output features from the second to sixth Conv-Blocks of
all three encoders are concatenated to feed into the decoder block via
attention units.
3.2.3. Decoder block
The proposed architecture leverages rich information extracted from
three types of inputs, i.e., 3D sub-volume consisting of 1 mm thick
CT slices, and forward and backward MIP images. This information
is extracted from encoders at multiple scales by skip connections to
feed into the decoder block. In the decoder block, we exploit the
attention gates (Schlemper et al., 2019) to enable the network to
focus on the expected nodule regions in coarse features coming from
the encoder block while discarding the redundant ones. The attention
units achieve this goal by performing element-wise multiplication of
input feature-maps and attention coefficients to highlight important
features as described in Fig. 6. Overall, the decoder block consists
of six deconvolution (De-Conv) blocks that up-samples the inputted
features while extracting useful information at different levels. Fig. 5(b)
describes the architecture of the De-Conv block, which consists of two
stages. In the first stage, two convolution blocks are followed by a batch
normalization layer. Input and output features of the first stage are
combined using a skip connection and forwarded to stage two. The
second stage is similar to the first one, followed by the up-sampling
layer. The decoder upscales the dimension of features step by step till
the final deconvolution block, which has the same feature dimension as
the ground truth mask, i.e., 256 × 256. Further, we extract the outputs
of each De-Conv block to feed forward to the distillation block, which
utilizes these features to reconstruct the nodule masks.
3.3. Self-distillation mechanism
The proposed framework employs a self-distillation mechanism to
improve the training process by using a self-distillation block, which
M. Usman et al.
Fig. 3. The illustration of 3 mm thick backward and forward MIP images with a corresponding normalized slice for four selected nodules. Each row, i.e., (a), (b), (c), and (d),
represents one nodule sample.
has been proven effective in optimizing the network’s performance for
classification tasks (Zhang et al., 2019). We extract the features from
four intermediate De-Conv blocks of the decoder block to pass into the
self-distillation block. Self-distillation block consists of four auxiliary
detectors that distil the knowledge from the main detector. The main
detector is connected with the deepest De-Conv block, therefore, having
access to the most refined features that provide enriched insights
for lung nodule detection. Each auxiliary detector block shares the
same architecture which is demonstrated in Fig. 5(c). The auxiliary
detector first immediately up-scales the features extracted from the
decoder block to the same dimensions as the inputs’ height and width,
i.e., 256 × 256. The second layer is a 2D convolution layer with Relu
activation, and the final layer is a convolution layer of a single filter
followed by softmax activation. Similarly, the main detector architec-
ture, which is shown in Fig. 5(d), has a 2D convolution layer with Relu
activation and a convolution layer of a single filter followed by softmax
activation.
During the training process, the main detector is trained with a
supervised training scheme by utilizing the ground truth, whereas all
the auxiliary detectors are trained as student models via distillation
from the main detector. Here, the main detector can be conceptually
regarded as the teacher model. To improve the learning ability of
6
Engineering Applications of Artificial Intelligence 129 (2024) 107597
auxiliary or student detectors in the proposed framework, we utilize
three types of losses during the training processes:
Loss 1: The dice loss that is calculated with labels of training data
to all the detector networks outputs. It provides equal opportunity to
each detector network to leverage the supervision from actual labels
for optimizing their weights.
Loss 2: KL (Kullback–Leibler) divergence loss under the main detector’s
supervision, which acts as a teacher network. We utilize the outputs of
each auxiliary detector, i.e., softmax output, and output of the main
detector to compute the KL divergence. This loss helps the distillation
of knowledge from the main detector to the auxiliary detectors, forcing
them to learn to produce similar results as the deepest detector.
Loss 3: This is L2 loss from hint (Adriana et al., 2015) that is computed
between the feature maps of the main detector network and each
auxiliary detector. The loss enables the auxiliary detectors to distill
the implicit knowledge in feature maps from the main detector, which
induces the auxiliary detectors’ feature maps to fit the feature maps of
the main detector.
3.4. False positive reduction using auxiliary detectors
In contrast to previous studies (e.g., Pezeshk et al. (2018) and Yuan
et al. (2021)) on dual-stage techniques, our framework exploits the
M. Usman et al. Engineering Applications of Artificial Intelligence 129 (2024) 107597

Fig. 4. Description of 3D dense block which transforms the 3D sub-volume of normalized thoracic CT scans into three-channel latent representation.

Fig. 6. The architecture of attention unit used in residual connections from encoders to
decoder block. Gating signal and feature maps come from the previous de-conv block
and the concatenated encoder features, respectively.

extracted from the 3D outputs of all four auxiliary detectors to detect

the false positives. Given that 𝑛 is the number of voxels in 𝑅𝑜𝑃𝑖 in
the 𝑖𝑡ℎ detector, we can define true positive determination criteria as
Fig. 5. The architectural details of convolutional, de-convolutional, and detector blocks
follows:
are illustrated in (a), (b), and (c), respectively. The architecture of the attention unit is
( )
1 ∑∑
used in residual connections from encoders to decoder blocks. Gating signal and feature 𝑘 𝑛
maps come from the previous de-conv block and the concatenated encoder features, 𝑖𝑠𝑇 𝑃 = 𝑡ℎ𝑟 𝑅𝑜𝑃𝑖 (𝑗), 𝜏 (3)
respectively. 𝑛𝑘 𝑖=1 𝑗=1
{
1, 𝜃 > 𝜏
𝑡ℎ𝑟(𝜃, 𝜏) = (4)
0, otherwise
auxiliary detectors of the same network to reduce the false positives at
the time of inference. All the nodules detected by the main detector are In the above equations, 𝑘 is the number of auxiliary detectors, which
considered nodule candidates, and the auxiliary detectors complement in our case are four, and 𝜏 is the threshold of the nodule and non-
the detection of the main detector to validate each nodule candidate. nodule probabilities. While 𝑡ℎ𝑟(.) is a binarized function to binarize the
Concretely, our main detector provides all the nodule candidates, and possibility value 𝜃, which is in the range of 0 to 1. We normalize the
we determine the 3D bounding box, by incorporating all the nodular whole submissions of probabilities by 𝑛𝑘 to limit the accumulation to
voxels, which act as a region of the proposal (𝑅𝑜𝑃 ) for each nodule 1. We apply the same criteria to all the nodule candidates after the
candidate. The similar 𝑅𝑜𝑃 𝑠 of the same dimensions and positions are inference to remove the false positives.

7
M. Usman et al.
Fig. 7. Description of the steps involved in the lung parenchyma segmentation. (a) Raw thoracic CT image. (b) Mask of the air regions. (c) Mask after the removal of irrelevant
objects other than the lung region. (d) Dilated mask of the lung parenchyma. (e) Image of segmented lung parenchyma.
4. Experimental setup
4.1. Datasets
In this work, we utilize the lung image database consortium and
image database resource initiative (LIDC/IDRI) dataset (Armato et al.,
2011) for all the experimentation. LIDC/IDRI dataset consists of 1018
thoracic computed tomography (CT) scans acquired from 1010 differ-
ent patients, which are annotated to facilitate computer-aided systems
for the assessment of lung nodule detection, classification and quantifi-
cation.The resolution of CT scans becomes particularly crucial when
the primary goal is the detection of lung nodules, which often re-
quire a higher resolution to distinguish and accurately classify. In the
LIDC/IDRI dataset, slice thickness varies from 0.6 mm to 5.0 mm. How-
ever, the scans with a slice thickness greater than 2.5 mm are typically
not recommended for lung nodule detection (Kazerooni et al., 2014).
Such scans might fail to capture the minute details of smaller nodules,
potentially leading to false negatives. Subsequently, we remove the
scans with slice thickness greater than 2.5 mm as such low-resolution
scans are not recommended for nodule screening (Kazerooni et al.,
2014).
However, while refining our dataset, we are also cognizant of
the fact that in real-world settings, we often encounter noisy data,
which may not always adhere to ideal conditions. It is essential for
a computer-aided detection system to be robust to such variabilities.
By focusing our study on high-resolution scans, we aim to build a
foundation upon which the system’s capabilities can be honed. Future
iterations and validations of the model could consider noisy and less-
than-ideal scans to gauge and further improve the system’s robustness.
LIDC/IDRI dataset is annotated in two steps by four radiologists.
Firstly, each radiologist annotated the nodules individually, and later
all the radiologists jointly analyzed each annotation to reassess their
annotations. For this study, we only utilized nodules having diameters
equal to or greater than 3 mm due to their clinical significance (Team,
2011). In order to avoid possible outliers, we only considered the
nodules accepted by at least 3 out of 4 radiologists as the reference
standard. After applying the aforementioned criteria, 888 scans were
left, which were utilized for all the experimentation.
4.2. Data preprocessing
We preprocess all the scans prior to feeding them into the proposed
MEDS-Net. Our preprocessing consists of two stages; lung parenchyma
segmentation and scan normalization.
In the first stage, we automatically segment the lung parenchyma
to exclude irrelevant regions such as clothes, machine objects, tissues,
spines, or ribs in the scans. Fig. 7 demonstrates the steps followed to
perform lung segmentation. Initially, an air mask is created, and wa-
tershed algorithm (Angulo and Jeulin, 2007) is applied to remove the
region other than the lung as shown in Fig. 7(b) and (c), respectively.
In order to avoid missing wall-attached lesions that might be nodules,
we keep more boundary information by binary morphology operations,
i.e., closing and dilation. Finally, the binarized scan was masked with
the original scan.
8
Engineering Applications of Artificial Intelligence 129 (2024) 107597
In the second stage, we perform the scan normalization. Since the
data set is collected from various CT scanners in the LIDC/IDRI dataset,
we set the window level from −1000 HU to 400 HU and normalized
images to the range between 0 and 1. We also normalized the slice
thickness to 1 mm and cropped each slice equally from all sides while
keeping the lung at the center for normalizing to 256 × 256 dimensions.
4.3. Training strategy
We divide the whole dataset scans into ten subsets to perform 10-
fold cross-validation. For each fold, training, validation, and test set
split are set to 80%, 10%, and 10% of the total scans, respectively. The
batch size is set to 3 due to the memory limitations of the GPU. We
used adam optimizer with an initial learning rate of 0.001 and first
and second momentum of 0.9 for the decay of the learning rate. We
use early stopping with patience of 10 epochs to avoid overfitting.
This study exploits three types of losses to train the model as
described in Section 3.3, which can be written as
Total = 1 + 2 + 3 (5)
To formulate these losses, we denote our main and auxiliary detec-
tors by 𝜃𝑚 and 𝜃𝑖∕𝑘 detectors, respectively. Where, 𝑖 ∈ [[1, 𝑘]] and 𝑘 is
the number of auxiliary detectors.
∑
𝑘
1 = DSC (𝜃𝑚 , 𝑦) + (1 − 𝛼) DSC (𝜃𝑖∕𝑘 , 𝑦) (6)
𝑖=1
2. ∣ 𝑥 ∩ 𝑦 ∣
DSC (𝑥, 𝑦) = 1 − (7)
∣𝑥 ∪ 𝑦∣
The second loss Kullback–Leibler (𝐾𝐿) divergence between each
auxiliary and main detector can be written as:
∑
𝑘
2 = 𝛼. 𝐾𝐿(𝜃𝑖∕𝑘 , 𝜃𝑚 ) (8)
𝑖=1
The third supervision comes from the features of the main deepest
detector to each auxiliary detector. Due to the difference in the size
of features inputted to each detector, additional convolution layers
in detector blocks are employed to align the dimensions prior to the
extraction of feature maps to calculate the feature loss. Feature loss
can be formulated as:
∑
𝑘
3 = 𝜆. ∣ 𝐹𝑖∕𝑘 , 𝐹𝑚 ∣ (9)
𝑖=1
where 𝐹𝑖∕𝑘 and 𝐹𝑚 represents the features from 𝜃𝑖∕𝑘 and 𝜃𝑚 detectors.
While two hyper-parameters 𝛼 and 𝜆 are used to moderate the role of
each loss.
4.4. Evaluation parameters
While comparing different CADe systems for lung nodule detection,
its imperative to focus on metrics that not only quantify performance
but also signify clinical relevance. In this section, we elaborate on the
evaluation metrics used throughout our manuscript to quantify the
performance of our lung nodule detection system. We also elucidate
M. Usman et al.
the importance of our chosen evaluation metrics in the context of the
proposed MEDS-Net’s efficacy.
4.4.1. Competition Performance Metric (CPM)
The Competition Performance Metric (CPM) is a measure specifi-
cally devised for comparing different computer-aided detection systems
in a competitive setting. Mathematically, the CPM for a system can be
represented as:
∑𝑁
𝑆𝑖
𝐶𝑃 𝑀 = 𝑖=1
𝑁 Our method
where 𝑆𝑖 is the sensitivity of the system for the 𝑖th test case and 𝑁
is the total number of test cases. A higher CPM indicates a superior
performance, denoting that the detection system is both accurate and
efficient.
4.4.2. Lung nodule detection sensitivity (%)
Sensitivity, often referred to as the True Positive Rate, quantifies
how effectively a system identifies the presence of lung nodules. It is
calculated using the formula:
𝑇𝑃
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦(%) = × 100
𝑇𝑃 + 𝐹𝑁
where 𝑇 𝑃 is the number of true positives (actual nodules correctly
identified) and 𝐹 𝑁 is the number of false negatives (actual nodules
missed by the system). Sensitivity offers a direct measure of the sys-
tem’s capability to correctly identify nodules, an essential trait for any
clinical detection system.
4.4.3. Average number of nodule candidates per scan
This metric gives an average count of potential nodule candidates
identified by the system for each scan. While not a direct measure of
accuracy, it provides insight into the system’s tendency towards being
conservative or aggressive in its detection approach. It is given by:
∑𝑁
𝐶𝑖
𝐴𝑣𝑒𝑟𝑎𝑔𝑒 𝐶𝑎𝑛𝑑𝑖𝑑𝑎𝑡𝑒𝑠 𝑝𝑒𝑟 𝑆𝑐𝑎𝑛 = 𝑖=1
𝑁 tion systems. Table 2 summarizes the results of candidate detection
where 𝐶𝑖 represents the number of candidates detected for the 𝑖th scan
and 𝑁 is the total number of scans.
4.4.4. False positives per scan
This metric measures the average number of false positive detec-
tions per scan, providing an indication of the system’s propensity for
making erroneous identifications. It can be represented as:
∑𝑁
𝐹 𝑃𝑖
𝐹 𝑃 𝑝𝑒𝑟 𝑆𝑐𝑎𝑛 = 𝑖=1
𝑁 exploiting the unidirectional MIP images of various thicknesses which
where 𝐹 𝑃𝑖 is the number of false positives for the 𝑖th scan and 𝑁 is the
total number of scans. A lower value for this metric is desirable as it
implies fewer incorrect nodule identifications per scan.
The Average Number of Nodule Candidates per Scan and False Positives
per Scan jointly provide insight into the system’s precision and potential
for over-detection. A balanced value indicates that MEDS-Net can effec-
tively differentiate nodules from non-nodular structures without being
overly conservative or aggressive.
5. Experimental results and discussion
In this section, we perform a thorough examination of the proposed
MEDS-Net framework, focusing on both its overall performance and the
findings from our ablation study. We begin in Section 5.1 by juxtapos-
ing the performance of our MEDS-Net with that of previously published
studies. Specifically, Section 5.1.1 evaluates our method for detecting
lung nodule candidates against existing research, while Section 5.1.2
extends the comparison to assess our framework following the false
positive reduction phase. Transitioning to Section 5.2, we direct our
attention to a comprehensive ablation study. Here, the efficacy of every
individual input and module within MEDS-Net is scrutinized. While
Section 5.2.1 narrows its focus to the nodule candidates detection stage,
Section 5.2.2 broadens the lens to examine the entirety of the MEDS-Net
framework.
9
Engineering Applications of Artificial Intelligence 129 (2024) 107597
Table 2
Performance comparison with other computer-aided detection systems on the LIDC/IDRI
dataset at nodule candidate detection stage.
CADe systems Sensitivity Total number Average number of
(%) of candidates candidates per scan
Zheng et al. (2019) 95.40 18,116 20.40
Pereira et al. (2021) 98.15 44,111 49.67
Wang et al. (2019) 96.80 53,484 60.20
Setio et al. (2017) 98.30 754,975 850.2
Zhao et al. (2022) 94.70 18,116 14.10
97.80 19,190 21.61
Note: The results presented are sourced directly from previously published research
papers.
5.1. Overall performance analysis
Typical CADe systems consist of two stages, i.e., candidate detection
and false positive reduction, and mostly two separate networks are
trained to achieve this goal. Nevertheless, in this study, the proposed
framework combines both stages by employing self-distillation during
the training and utilizing the auxiliary detectors for false positive reduc-
tion. For a comprehensive analysis of the proposed CADe system, we
compare our performance with previously published methods for the
candidate detection stage as well as after the false positive reduction
stage. Below we compare the performance of the proposed MEDS-Net
with multiple state-of-the-art studies.
5.1.1. Performance at nodule candidates detection stage
All the nodules detected by the main, the deepest detector, of the
proposed MEDS-Net are considered nodule candidates. The capability
of nodule detection of the proposed framework has been compared with
the recently published deep learning-based individual candidate detec-
stages of previously published works. Pereira et al. (2021) and Setio
et al. (2017) outperform the proposed method in terms of sensitivity.
However, in order to localize as many candidates as possible, these
CADe systems obtained a several-fold higher number of FPs. Although
Zheng et al. (2019) and Zhao et al. (2022) achieved the lowest can-
didates per scan, these CADe systems had lower sensitivity compared
to our proposed scheme. Most importantly, among these two CADe
systems, Zheng et al. (2019) obtained superior sensitivity of 95% by
shows the efficacy of MIP image for lung nodule detection. Zheng et al.
(2019) trained four individual 2D networks (i.e., UNet) with unidirec-
tional MIP images to obtain such higher sensitivity. Nevertheless, our
framework achieved 97.8% sensitivity with a single self-distillation-
based network by leveraging the bidirectional MIP images. The results
demonstrate that incorporating bidirectional MIP images in the pro-
posed self-distilled network significantly improves the performance on
nodule detection.
5.1.2. Performance after false positives reduction
To provide a comparative analysis of the complete pipeline of the
proposed framework, we chose recent DL-based studies which uti-
lized the LIDC/IDRI dataset. We select competition performance metric
(CPM) (Niemeijer et al., 2010) for comparison. Table 3 summarizes the
performance of these techniques. In the table, the columns under ’False
Positive Per Scan’ represent the sensitivity of the CADe systems at vari-
ous allowed false positive rates. Specifically, for each false positive rate
(e.g., 0.5 false positives per scan), the corresponding value illustrates
the proportion of actual nodules the system could detect. Therefore,
a higher value indicates superior nodule detection capability for that
specific false positive allowance. Most of the listed methods are dual-
stage in which nodule candidate detection and false positive reduction
are performed with different networks. Among all the listed methods,
our proposed method outperforms in terms of CPM which demonstrates
M. Usman et al. Engineering Applications of Artificial Intelligence 129 (2024) 107597

Table 3
Complete pipeline performance comparison with other computer-aided detection systems on the LIDC/IDRI dataset.
CADe system Year Type of scheme False positive per scan CPM
0.125 0.25 0.5 1 2 4 8
Wang et al. (2019) 2018 Dual-stage 0.788 0.847 0.895 0.934 0.952 0.959 0.963 0.903
Pezeshk et al. (2018) 2019 Dual-stage 0.637 0.723 0.804 0.865 0.907 0.938 0.952 0.832
Zheng et al. (2019) 2019 Dual-stage 0.876 0.899 0.912 0.927 0.942 0.948 0.953 0.922
Cao et al. (2020) 2020 Dual-stage 0.848 0.899 0.925 0.936 0.949 0.957 0.96 0.925
Li and Fan (2020) 2020 Single-stage 0.60 0.674 0.751 0.824 0.85 0.853 0.859 0.773
Pereira et al. (2021) 2021 Dual-stage 0.812 0.875 0.918 0.949 0.957 0.970 0.976 0.922
Zhou et al. (2022) 2022 Dual-stage 0.742 0.840 0.899 0.925 0.944 0.954 0.959 0.895
Zhao et al. (2022) 2022 Dual-stage 0.656 0.754 0.833 0.917 0.951 0.97 0.977 0.865
Zhu et al. (2022) 2022 Single-stage 0.782 0.834 0.893 0.917 0.932 0.952 0.956 0.895
Mei et al. (2021) 2022 Dual-stage 0.712 0.802 0.865 0.901 0.937 0.946 0.955 0.874
Guo et al. (2021) 2022 Single-stage 0.832 0.886 0.928 0.937 0.946 0.952 0.959 0.92
Our framework 2023 Single-stage 0.883 0.915 0.928 0.941 0.953 0.962 0.968 0.936

Table 4
Summary of results obtained from different configurations of proposed model on LIDC/IDRI dataset for lung nodule detection stage.
Model # Input type Attention Self distillation Performance
3D Sub-volume Forward MIP images Backward MIP images Sensitivity (%) Total number of candidates Candidates per scan
1 ✓ ✗ ✗ 90.22 42,476 47.83
2 ✗ ✓ ✗ 93.84 36,281 40.86
3 ✗ ✗ ✓ ✓ ✓ 94.10 35,958 40.49
4 ✓ ✓ ✗ 96.12 24,345 27.42
5 ✗ ✓ ✓ 95.78 27,863 31.38
6 ✗ ✗ 92.16 41,954 47.25
7 ✓ ✗ 93.59 30,096 33.89
✓ ✓ ✓
8 ✗ ✓ 95.45 22,847 25.73
9 ✓ ✓ 97.81 19,190 21.61

the evident effectiveness of the proposed auxiliary detectors-based false
positive reduction scheme. Pereira et al. (2021) had slightly better
performance for higher FPs per scan (i.e., ≥1). A possible explanation is
that they had utilized a two-stage scheme in which they first exploited
a mask region-convolutional neural network (Mask R-CNN) for nodule
detection and employed a classifier ensemble based on CT attenuation
patterns for false positive reduction, making the technique more com-
plex. Nevertheless, our method performs significantly better for smaller
values of FPs per scan (i.e., <1). Similarly, Zheng et al. (2019) obtained
better performance than Pereira et al. (2021) for smaller values of FPs
per scan (i.e., <1) which can be attributed to the incorporation of MIP
images which enhances the nodule detection capability of networks.
Zheng et al. (2019) utilized unidirectional MIP images, whereas our
scheme exploits bidirectional MIPs with different levels of thickness to
achieve superior results. Most importantly, among single-stage meth-
ods, our technique achieved significantly better performance for all the
range of FPs per scan which proves that incorporating self-distillation
greatly assists the optimization of the network to improve the lung
detection rate efficiently.
5.2. Ablation study
To demonstrate the effectiveness of each component incorporated
into the proposed MEDS-Net architecture, we implemented various ver-
sions of the proposed framework. We analyzed results after the nodule
detection stage as well as after the false positives reduction stage. The
following subsections describe the details of these experiments and the
corresponding results.
5.2.1. Analysis of candidates detection stage
We initiate our evaluation by analyzing the contribution of each
component of MEDS-Net for the lung nodule detection stage. For this
purpose, we implemented eight downgraded versions of MEDS-Net,
having different configurations and architecture, to evaluate their lung
nodules detection performance. Table 4 summarizes the results ob-
tained from these models for the lung nodules detection stage. First, to
fairly analyze the effect of each input individually, models 1 to 3 utilize
single encoder-based architecture which takes only one input among 3D
sub-volume, forward, and backward MIP images, respectively. These
networks are equipped with attention units and self-distillation mech-
anisms for lung nodule detection tasks. The results demonstrate that
models 2 and 3 with MIP images of various thicknesses achieved better
sensitivity with a lower number of candidates than the 3D input-based
model. This proves that the incorporation of MIP images significantly
improves the learning ability of networks which helps to better dis-
tinguish between nodule and non-nodular structures. Whereas, models
2 and 3 depict quite similar performance which shows that forward
and backward MIP images provide different yet the same quantity
of information, i.e., unidirectional insights, to the network and their
individual effect is the same. Models 4 and 5 are trained with two types
of inputs by using dual-encoder-based architecture having attention
units and a self-distillation mechanism. Particularly, we combined 3D
sub-volume with forward MIP images and forward MIP images with
backward MIP images as inputs to models 4 and 5, respectively. It
can be observed that combining the two types of inputs improves
the sensitivity and also reduces the number of candidates per scan.
However, model 4 performed slightly better than model 5 which can
be attributed to the diverse type of inputs that enable the network
to leverage the exceptional abilities of MIP images to provide deeper
insights and 3D patch input that improves the localization of nodules.
The visual examination of feature maps offers a deeper understand-
ing of performance differences, and it is paramount to appreciate the
advantages brought by the inclusion of forward and backward MIP
images in our proposed MEDS-Net. To elucidate this, we delve into the
Class-Activation Map (CAM) from models 1, 4, and 9, corresponding
to single, dual, and triple input configurations, detailed in Table 4.
Refer to Fig. 9, which contrasts the CAMs for the said models against
the ground truth. Model 1, relying solely on raw 3-D slices, manifests
more diffuse feature maps, inadvertently drawing attention to non-
nodular regions, leading to increased false positives. In contrast, model
4, assimilating forward MIP images and 3-D raw slices, presents more
concentrated feature maps. Yet, it occasionally struggles to discern
bronchi and vessels from genuine nodules. Model 9, the cornerstone of
our proposal—the MEDS-Net, takes in a trio of inputs: raw 3-D CT slices
and both forward and backward MIP images. Evidently, it exhibits the
most adeptness, underscoring that the bidirectional MIP images amplify

10
M. Usman et al.
Fig. 8. Illustration of detected nodule candidates by all four auxiliary detectors and
main detector on three samples, i.e., (a), (b), and (c). Bounding boxes in green,
yellow, and red colors represent the gold standard, true positives, and false positives,
respectively.
the MEDS-Net’s efficacy in distinguishing nodules from non-nodular
areas.
From models 6 to 9, we analyze the effectiveness of incorporat-
ing the attention units and self-distillation in our proposed archi-
tecture. For this purpose, we fixed our input pipeline and utilized
multi-encoder-based architecture incorporating the 3D sub-volume and
bidirectional (forward and backward) MIP images. Among the four
versions, in model 6, we observe severe degradation in the sensi-
tivity when attention units and self-distillation are removed. It also
immensely increased the number of nodule candidates, which shows
that the attention and self-distillation mechanism significantly con-
tributes to reducing the number of false positives at the candidate
detection stage. Concretely, the individual incorporation of attention
units and self-distillation mechanism improve 3.94% and 5.78% sen-
sitivity, respectively. Most importantly, the self-distillation mechanism
greatly reduces the number of candidates, improving the architecture’s
11
Engineering Applications of Artificial Intelligence 129 (2024) 107597
efficiency. Finally, the best performance is shown by model 9, which
represents the proposed MEDS-Net architecture, having attention and
a self-distillation mechanism with three types of inputs. The results
evidence that the combination of attention and self-distillation mecha-
nism greatly increases our architecture’s learning ability, enabling it to
better distinguish between the nodules and non-nodular structures in
the complex lung region.
To analyze the impact of self-distillation, we implemented the vari-
ant of the proposed architecture without self-distillation. Since MEDS-
Net is a deeper network, it is critical to ensure that the network
is optimally trained without encountering vanishing gradient prob-
lems (Glorot and Bengio, 2010). Subsequently, we analyzed the effect of
self-distillation during the training process. The training and validation
curves from the deepest detector of both variants have been shown in
Fig. 10. Conceptually, self-distillation is an extended version of deep
supervision that is proven to be effective in addressing the vanish-
ing gradient problem for segmentation networks (Dou et al., 2016).
The results show that the self-distillation mechanism in the proposed
framework assists the training process to optimize network learning.
It improves the convergence rate and helps optimize the learning by
achieving lower training and validation losses. These improvements can
be attributed to the supervision of the intermediate layers at multiple
levels, which encourages the earlier layers to learn more meaningful
and representative features for improved detection of lung nodules.
We further analyze the performance of each detector, including
main and auxiliary detectors, in the proposed MEDS-Net. Table 5
summarized the results obtained from all the detectors, including the
baseline implementation of multi-encoder-based architecture without
self-distillation. An ensemble result is obtained by including all the
nodule candidates detected by each detector. The results of those
detectors with inferior performance to the baseline model have been
highlighted in red. Further, the qualitative results of each detector have
been illustrated in Fig. 8 on three different slices containing two or
more nodules. The results demonstrate that (i) The proposed MEDS-
Net has effectively detected almost all the nodules and achieved the
sensitivity 99.41% with its five detectors. (ii) The overall performance
of deeper detectors is improved; however, several nodules missed by
deeper networks have been detected by the less deep network, indicat-
ing that each detector has detected nodules independently. The same
has been demonstrated in Fig. 8. (iii) The self-distillation plays a crucial
role in improving the deepest network, the main detector. Introducing
four auxiliary detectors improves the 4.08% sensitivity of the main
detector and significantly reduces the number of false positives per
scan.
5.2.2. Analysis after false positive reduction
In this section, we analyze the effect of each input incorporated in
the proposed MEDS-Net architecture on the complete pipeline. Con-
cretely, the MEDS-Net exploits bidirectional MIP images along with
the 3D patch to detect the lung nodules; therefore, it is important
to analyze the impact of bidirectional MIP images after false positive
reduction. We implemented single-encoder and dual-encoder versions
of the proposed architecture with a complete pipeline that includes the
false positive reduction stage. We implemented three single-encoder-
based variants that take one input at a time and trained these models
with the 3D patch, forward, and backward image inputs individually.
Similarly, we implemented two dual-encoder-based variants, i.e., one
with bidirectional MIP image inputs and another with 3D patch and
forward MIP images. All the variants are implemented with a self-
distillation mechanism and followed by the same pipeline. To compare
the performances of these variants models with the proposed MEDS-Net
for nodule detection, we analyze the free-response receiver operating
characteristic (FROC) (Bandos et al., 2009) curves of each variant
shown in Fig. 11. The single-encoder-based architecture with 3D patch
input has the lowest performance, and it becomes worse when we
reduce the number of false positives per scan. Since the 3D patch of
M. Usman et al. Engineering Applications of Artificial Intelligence 129 (2024) 107597

Fig. 9. Visualization of CAMs for diverse input configurations in the MEDS-Net. Each row showcases an input image demarcated by the ground truth, succeeded by CAMs from
models #1, #4, and #9.

Fig. 10. Comparison of learning curves of the training and validation curves of proposed multi-encoder based networks with and without self-distillation mechanism has been
demonstrated.

scan consists of 11 slices of 1 mm thickness (i.e., of depth 10 mm)
which is insufficient to extract detailed insights of the nodule in the
complex surroundings within the lung, it is challenging to detect nod-
ules without false positives. The nodule bigger than 10 mm appear
similar to the vessels, and small vessels have quite similar structural
appearance to the nodule, which confuses the network and results in
poor performance. Since conceptually, both directions of MIP images
contain the same amount of information, subsequently, the single-
encoder versions with forward and backward MIP images demonstrate
similar; however, better performance compared to its 3D patch variant.
MIP image based models are able to achieve almost 80% sensitivity
even with 0.125 FPs per scan which demonstrates the effectiveness of
MIP images for lung nodule detection.
We also compared dual-encoder-based architectures, for which we
implemented two combinations. At first, we utilized bidirectional MIP
images and achieved significantly improved performance, which can
be attributed to the fact that MIP images enhance the visibility of
lung nodules, making it easier for the network to distinguish them

12
M. Usman et al. Engineering Applications of Artificial Intelligence 129 (2024) 107597

Fig. 11. Free-response receiver operating characteristic (FROC) curves of various versions of proposed MEDS-Net architecture along with MEDS-Net results on the LIDC/IDRI
database.

Table 5
Performance comparison of each detector of proposed MEDS-Net for lung nodule detection on different size of nodules.
Total number of scans: 888
Total number of nodules: 1186 (3–10 mm: 905, 10–20 mm: 231, ≥20 mm: 50)
Feature level No. of detected No. of detected No. of detected Total No. of Sensitivity False positives FPs per scan
nodules (3–10 nodules (10–20 nodules (≥20 detected nodules (%) (FPs)
mm) mm) mm)
Multi-encoder net 862 202 50 1114 93.93 28 910 32.56
(without
self-distillation)
Auxiliary detector (1/4) 841 195 46 1082 91.23 30 226 34.04
Auxiliary detector (2/4) 857 198 48 1103 93 26 954 30.35
Auxiliary detector (3/4) 856 205 50 1111 93.68 22 047 24.83
Auxiliary detector (4/4) 877 212 50 1139 96.04 19 338 21.78
Main detector 889 221 50 1160 97.81 18 004 20.27
Ensemble 901 228 50 1179 99.41 36 780 41.42

from other anatomical structures. Most importantly, bidirectional MIP
images cover a reasonable depth of scan, i.e., 30 mm, which enables the
network to better explore the insights of each suspected nodule region.
However, MIP images have certain limitations while representing the
information as they rely on maximum intensities, suppressing the infor-
mation in the intermediate intensities. Therefore, in clinical practice,
radiologists examine the nodules in the raw CT slices after the initial
screen in MIP images. To this end, we also implemented a dual-encoder-
based variant with a 3D patch and forward MIP images. The results
demonstrate that the combination of 1 mm thick stack of slices with
unidirectional MIP images are more effective than bidirectional MIPs
as it equips the network with rich information which contains high-
level 3D aspect in the form of MIP as well as low-level, 2D slice-level
information that helps to capture the subtle variations and precisely
locate the small nodules. Finally, combining raw 3D patches of scan
with bidirectional MIPs images in MEDS-Net depicts the optimal and
consistent performance for lung nodule detection.
To assess the effectiveness of the proposed auxiliary false posi-
tive reduction technique, we implemented a dual-stage variant of our
framework. In this scheme, MEDS-Net remains consistent for detecting
nodule candidates as in the single-stage approach. However, for false
positive reduction, two distinct 3-D CNNs are used those are fed with
volume sizes of 32 × 32 × 32 and 16 × 16 × 16 voxels, analogous to
the method in Zheng et al. (2019). All nodule candidates identified
by MEDS-Net are used to train both 3-D CNNs, which then compute
the likelihood of a nodule being present. Table 6 summarizes the
performance metrics: false positive reduction (evaluated via CPM score)
and computational efficiency (measured in terms of the number of
networks involved, the total count of network parameters, cumulative
training time, and overall inference duration). The findings suggest that
our proposed technique not only excels over traditional false positive
reduction methods but also showcases computational efficiency. The
shorter training and inference times are especially crucial for real-
time clinical scenarios. The noteworthy performance in reducing false
positives can be attributed to the enhanced learning capability of the
auxiliary branches in MEDS-Net, a result of shared weights, which
bolsters nodule recognition. Given that our method requires no addi-
tional networks, only one network is trained, leading to a reduction
in the number of network parameters. This translates to less time
being needed for both training and inference, emphasizing the proposed
scheme’s suitability for clinical implementation.
6. Conclusions and future works
This work proposes a novel multi-encoder-based self-distillation
network (MEDS-Net) that leverages bidirectional maximum intensity
projection (MIP) images for a lung nodule detection system. Partic-
ularly, along with a 3D patch of the scan, the proposed framework
utilized forward and backward MIP images of 3, 5, and 10 mm thick-
ness as input to improve the network learning. Most importantly, unlike

13
M. Usman et al.
Table 6
Performance and computational comparison of single-stage and dual-stage versions of proposed MEDS-Net for lung nodule
detection.
Framework details CPM Total No. of
networks trained
MEDS-Net (Single-stage 0.936 1 510,566
architecture)
MEDS-Net (Dual-stage 0.928 3 723,301
architecture)
conventional computer-aided detection (CADe) systems, the study ex-
ploited the auxiliary detectors of the same network to reduce the false
positives, which avoids the extra computational costs caused by the
separate false positive reduction networks. The key highlights are as
follows:
1. The framework has been comprehensively evaluated using the
widely used LIDC/IDRI public dataset. The results using the
proposed MEDS-Net showed that this network has excellent
detection performance for diverse nodule morphology and has
the ability to distinguish false-positive nodules accurately. The
CPM scores of the proposed method can be as high as 0.936,
superior to the most advanced competitive networks.
2. MEDS-Net exploits multi-scale features learning by using in-
termediate auxiliary outputs branches originating from various
levels of decoder block to minimize the false positives.
3. The experiments have shown that employment of self-distillation
significantly improves the learning process by supervising the
intermediate layers of the decoder block. Ablation experiments
have also demonstrated that all the components in the pro-
posed MEDS-Net are chosen carefully for effective lung nodule
detection.
The exceptional nodule detection capability of our proposed CADe
system holds promise not only in assisting radiologists with swift lung
nodule diagnosis but also in laying the groundwork for a holistic,
automated lung diagnosis system capable of both nodule segmentation
and classification based on severity and texture of lung nodules.
Although the study effectively emulated clinical workflows via the
MEDS-Net architecture, a clear limitation is the framework’s inter-
pretability, a critical facet in the realm of medical imaging. Moreover,
the validation of our framework was predominantly anchored on the
benchmark LIDC/IDRI dataset. However, to ascertain its broad applica-
bility and robustness, its imperative to deploy and evaluate it in real-
time clinical scenarios. As a forward-looking measure, our subsequent
endeavors aim to bolster the framework’s interpretability, potentially
through algorithm unrolling techniques, and undertake comprehensive
evaluations in real-world clinical settings.
CRediT authorship contribution statement
Muhammad Usman: Conceptualization, Methodology, Software,
Validation, Formal analysis, Writing – original draft. Azka Rehman:
Software, Writing – original draft, Investigation, Validation. Abdullah
Shahid: Software, Investigation, Visualization. Siddique Latif: For-
mal analysis, Writing – review & editing. Yeong-Gil Shin: Validation,
Supervision.
Declaration of competing interest
The authors declare that they have no known competing finan-
cial interests or personal relationships that could have appeared to
influence the work reported in this paper.
Data availability
The authors do not have permission to share data.
14
Engineering Applications of Artificial Intelligence 129 (2024) 107597
Total number Total training Total inference
of parameters time (h) time (s)
3.416 25
4.666 38
References
Adriana, R., Nicolas, B., Ebrahimi, K.S., Antoine, C., Carlo, G., Yoshua, B., 2015. Fitnets:
Hints for thin deep nets. In: Proc. ICLR, Vol. 2.
Angulo, J., Jeulin, D., 2007. Stochastic watershed segmentation. In: ISMM (1). pp.
265–276.
Armato, III, S.G., McLennan, G., Bidaut, L., McNitt-Gray, M.F., Meyer, C.R.,
Reeves, A.P., Zhao, B., Aberle, D.R., Henschke, C.I., Hoffman, E.A., et al., 2011.
The lung image database consortium (LIDC) and image database resource initiative
(IDRI): a completed reference database of lung nodules on CT scans. Med. Phys.
38 (2), 915–931.
Baldwin, D.R., 2015. Prediction of risk of lung cancer in populations and in pulmonary
nodules: significant progress to drive changes in paradigms. pp. 1–3.
Bandos, A.I., Rockette, H.E., Song, T., Gur, D., 2009. Area under the free-response ROC
curve (FROC) and a related summary index. Biometrics 65 (1), 247–256.
Cao, H., Liu, H., Song, E., Ma, G., Xu, X., Jin, R., Liu, T., Hung, C.-C., 2020. A two-stage
convolutional neural networks for lung nodule detection. IEEE J. Biomed. Health
Inform. 24 (7), 2006–2015.
Dong, H., Chen, Z., Zhao, J., Yuan, M., Yu, F., Zhang, J., Zhang, L., 2022. Abdominal
organ segmentation via self training.
Dou, Q., Chen, H., Jin, Y., Yu, L., Qin, J., Heng, P.-A., 2016. 3D deeply supervised
network for automatic liver segmentation from ct volumes. In: International
Conference on Medical Image Computing and Computer-Assisted Intervention.
Springer, pp. 149–157.
Glorot, X., Bengio, Y., 2010. Understanding the difficulty of training deep feedforward
neural networks. In: Proceedings of the Thirteenth International Conference on
Artificial Intelligence and Statistics. JMLR Workshop and Conference Proceedings,
pp. 249–256.
Gruden, J.F., Ouanounou, S., Tigges, S., Norris, S.D., Klausner, T.S., 2002. Incremental
benefit of maximum-intensity-projection images on observer detection of small
pulmonary nodules revealed by multidetector CT. Am. J. Roentgenol. 179 (1),
149–157.
Guleryuz Kizil, P., Hekimoglu, K., Coskun, M., Akcay, S., 2020. Diagnostic importance
of maximum intensity projection technique in the identification of small pulmonary
nodules with computed tomography.
Guo, Z., Zhao, L., Yuan, J., Yu, H., 2021. MSANet: Multiscale aggregation network
integrating spatial and channel information for lung nodule detection. IEEE J.
Biomed. Health Inf. 26 (6), 2547–2558.
Halder, A., Dey, D., Sadhu, A.K., 2020. Lung nodule detection from feature engineering
to deep learning in thoracic CT images: a comprehensive review. J. Digit. Imaging
33 (3), 655–677.
He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep residual learning for image recog-
nition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition. pp. 770–778.
Hinton, G., Vinyals, O., Dean, J., 2015. Distilling the knowledge in a neural network.
arXiv preprint arXiv:1503.02531.
Hui, Z., Wang, X., Gao, X., 2018. Fast and accurate single image super-resolution
via information distillation network. In: Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition. pp. 723–731.
Jabeen, N., Qureshi, R., Sattar, A., Baloch, M., 2019. Diagnostic accuracy of maximum
intensity projection in diagnosis of malignant pulmonary nodules. Cureus 11 (11).
Jiang, L., Zhou, W., Li, H., 2019. Knowledge distillation with category-aware attention
and discriminant logit losses. In: 2019 IEEE International Conference on Multimedia
and Expo (ICME). IEEE, pp. 1792–1797.
Kazerooni, E.A., Austin, J.H., Black, W.C., Dyer, D.S., Hazelton, T.R., Leung, A.N.,
McNitt-Gray, M.F., Munden, R.F., Pipavath, S., 2014. ACR–STR practice parameter
for the performance and reporting of lung cancer screening thoracic computed
tomography (CT): 2014 (resolution 4). J. Thorac. Imaging 29 (5), 310–316.
Li, Y., Fan, Y., 2020. Deepseed: 3D squeeze-and-excitation encoder-decoder convo-
lutional neural networks for pulmonary nodule detection. In: 2020 IEEE 17th
International Symposium on Biomedical Imaging (ISBI). IEEE, pp. 1866–1869.
Li, Q., Jin, S., Yan, J., 2017. Mimicking very efficient network for object detection. In:
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.
pp. 6356–6364.
Liu, L., Ouyang, W., Wang, X., Fieguth, P., Chen, J., Liu, X., Pietikäinen, M., 2020. Deep
learning for generic object detection: A survey. Int. J. Comput. Vis. 128, 261–318.
Luan, Y., Zhao, H., Yang, Z., Dai, Y., 2019. Msd: Multi-self-distillation learning via
multi-classifiers within deep neural networks. arXiv preprint arXiv:1911.09418.
M. Usman et al.
Luo, X., Song, T., Wang, G., Chen, J., Chen, Y., Li, K., Metaxas, D.N., Zhang, S.,
2022. SCPM-net: An anchor-free 3D lung nodule detection network using sphere
representation and center points matching. Med. Image Anal. 75, 102287.
Masood, A., Sheng, B., Yang, P., Li, P., Li, H., Kim, J., Feng, D.D., 2020. Automated
decision support system for lung cancer detection and classification via enhanced
RFCN with multilayer fusion RPN. IEEE Trans. Ind. Inform. 16 (12), 7791–7801.
Mei, J., Cheng, M.-M., Xu, G., Wan, L.-R., Zhang, H., 2021. SANet: A slice-aware
network for pulmonary nodule detection. IEEE Trans. Pattern Anal. Mach. Intell..
Monkam, P., Qi, S., Ma, H., Gao, W., Yao, Y., Qian, W., 2019. Detection and
classification of pulmonary nodules using convolutional neural networks: a survey.
IEEE Access 7, 78075–78091.
Ni, Z.-L., Zhou, X.-H., Wang, G.-A., Yue, W.-Q., Li, Z., Bian, G.-B., Hou, Z.-G., 2022.
SurgiNet: Pyramid attention aggregation and class-wise self-distillation for surgical
instrument segmentation. Med. Image Anal. 76, 102310.
Niemeijer, M., Loog, M., Abramoff, M.D., Viergever, M.A., Prokop, M., van Ginneken, B.,
2010. On combining computer-aided detection systems. IEEE Trans. Med. Imaging
30 (2), 215–223.
Pereira, F.R., De Andrade, J.M.C., Escuissato, D.L., De Oliveira, L.F., 2021. Classifier
ensemble based on computed tomography attenuation patterns for computer-aided
detection system. IEEE Access 9, 123134–123145.
Pezeshk, A., Hamidian, S., Petrick, N., Sahiner, B., 2018. 3-D convolutional neural
networks for automatic detection of pulmonary nodules in chest CT. IEEE J.
Biomed. Health Inform. 23 (5), 2080–2090.
Phuong, M., Lampert, C.H., 2019. Distillation-based training for multi-exit architectures.
In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp.
1355–1364.
Qin, D., Bu, J.-J., Liu, Z., Shen, X., Zhou, S., Gu, J.-J., Wang, Z.-H., Wu, L., Dai, H.-F.,
2021. Efficient medical image segmentation based on knowledge distillation. IEEE
Trans. Med. Imaging 40 (12), 3820–3831.
Ronneberger, O., Fischer, P., Brox, T., 2015. U-net: Convolutional networks for
biomedical image segmentation. In: International Conference on Medical Image
Computing and Computer-Assisted Intervention. Springer, pp. 234–241.
Schlemper, J., Oktay, O., Schaap, M., Heinrich, M., Kainz, B., Glocker, B., Rueckert, D.,
2019. Attention gated networks: Learning to leverage salient regions in medical
images. Med. Image Anal. 53, 197–207.
Seemann, M., Staebler, A., Beinert, T., Dienemann, H., Obst, B., Matzko, M., Pis-
titsch, C., Reiser, M., 1999. Usefulness of morphological characteristics for the
differentiation of benign from malignant solitary pulmonary lesions using HRCT.
Eur. Radiol. 9 (3), 409–417.
Setio, A.A.A., Traverso, A., De Bel, T., Berens, M.S., Van Den Bogaard, C., Cerello, P.,
Chen, H., Dou, Q., Fantacci, M.E., Geurts, B., et al., 2017. Validation, comparison,
and combination of algorithms for automatic detection of pulmonary nodules in
computed tomography images: the LUNA16 challenge. Med. Image Anal. 42, 1–13.
Sluimer, I., Schilham, A., Prokop, M., Van Ginneken, B., 2006. Computer analysis of
computed tomography scans of the lung: a survey. IEEE Trans. Med. Imaging 25
(4), 385–405.
Sung, H., Ferlay, J., Siegel, R.L., Laversanne, M., Soerjomataram, I., Jemal, A.,
Bray, F., 2021. Global cancer statistics 2020: GLOBOCAN estimates of incidence
and mortality worldwide for 36 cancers in 185 countries. CA: Cancer J. Clin. 71
(3), 209–249.
15
Engineering Applications of Artificial Intelligence 129 (2024) 107597
Team, N.L.S.T.R., 2011. Reduced lung-cancer mortality with low-dose computed
tomographic screening. N. Engl. J. Med. 365 (5), 395–409.
Tung, F., Mori, G., 2019. Similarity-preserving knowledge distillation. In: Proceedings
of the IEEE/CVF International Conference on Computer Vision. pp. 1365–1374.
Wallis, J.W., Miller, T.R., Lerner, C.A., Kleerup, E.C., 1989. Three-dimensional display
in nuclear medicine. IEEE Transactions on Medical Imaging 8 (4), 297–230.
Wang, J., Wang, J., Wen, Y., Lu, H., Niu, T., Pan, J., Qian, D., 2019. Pulmonary nodule
detection in volumetric chest CT scans using CNNs-based nodule-size-adaptive
detection and classification. IEEE Access 7, 46033–46044.
Wang, X., Zheng, Y., Zhang, X., Cao, T., Li, C., Wang, X., Wang, Y., Fang, Z.,
Wang, Y., 2022. Self-knowledge distillation for the object segmentation based on
atrous spatial pyramid. In: Journal of Physics: Conference Series, Vol. 2294. IOP
Publishing, 012023, no. 1.
Wu, Z., Zhou, Q., Wang, F., 2021. Coarse-to-fine lung nodule segmentation in CT images
with image enhancement and dual-branch network. IEEE Access 9, 7255–7262.
Xie, L., Cai, W., Gao, Y., 2022. Dmcgnet: A novel network for medical image
segmentation with dense self-mimic and channel grouping mechanism. IEEE J.
Biomed. Health Inf. 26 (10), 5013–5024.
Xu, G., Liu, Z., Li, X., Loy, C.C., 2020a. Knowledge distillation meets self-supervision.
In: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August
23–28, 2020, Proceedings, Part IX. Springer, pp. 588–604.
Xu, K., Rui, L., Li, Y., Gu, L., 2020b. Feature normalized knowledge distillation for
image classification. In: Computer Vision–ECCV 2020: 16th European Conference,
Glasgow, UK, August 23–28, 2020, Proceedings, Part XXV 16. Springer, pp.
664–680.
Yuan, H., Wu, Y., Cheng, J., Fan, Z., Zeng, Z., 2021. Pulmonary nodule detection using
3-D residual U-net oriented context-guided attention and multi-branch classification
network. IEEE Access 10, 82–98.
Zhang, L., Song, J., Gao, A., Chen, J., Bao, C., Ma, K., 2019. Be your own teacher:
Improve the performance of convolutional neural networks via self distillation. In:
Proceedings of the IEEE/CVF International Conference on Computer Vision. pp.
3713–3722.
Zhang, J., Xia, Y., Cui, H., Zhang, Y., 2018. Pulmonary nodule detection in medical
images: A survey. Biomed. Signal Process. Control 43, 138–147.
Zhao, D., Liu, Y., Yin, H., Wang, Z., 2022. An attentive and adaptive 3D CNN for
automatic pulmonary nodule detection in CT image. Expert Syst. Appl. 118672.
Zheng, S., Cui, X., Vonder, M., Veldhuis, R.N., Ye, Z., Vliegenthart, R., Oudkerk, M.,
van Ooijen, P.M., 2020. Deep learning-based pulmonary nodule detection: Effect of
slab thickness in maximum intensity projections at the nodule candidate detection
stage. Comput. Methods Programs Biomed. 196, 105620.
Zheng, S., Guo, J., Cui, X., Veldhuis, R.N., Oudkerk, M., Van Ooijen, P.M., 2019.
Automatic pulmonary nodule detection in CT scans using convolutional neural
networks based on maximum intensity projection. IEEE Trans. Med. Imaging 39
(3), 797–805.
Zhou, Z., Gou, F., Tan, Y., Wu, J., 2022. A cascaded multi-stage framework for auto-
matic detection and segmentation of pulmonary nodules in developing countries.
IEEE J. Biomed. Health Inf..
Zhu, X., Wang, X., Shi, Y., Ren, S., Wang, W., 2022. Channel-wise attention mechanism
in the 3D convolutional network for lung nodule detection. Electronics 11 (10),
1600.