Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

False Positive Reduction of Pulmonary Nodule on CT image using

Attention-based Multiple Instance Learning


Chi Cuong Nguyen Giang Son Tran Vincent Nguyen
ICTLab, University of Science and ICTLab, University of Science and Université d’Orléans
Technology of Hanoi, Vietnam Technology of Hanoi, Vietnam Orléans, France
Academy of Science and Technology Academy of Science and Technology
of Hanoi of Hanoi
Hanoi, Vietnam Hanoi, Vietnam
ABSTRACT are significant for treatment and survival. However, the beginning
Diagnosis and treatment of multiple pulmonary nodules are clin- symptoms of lung cancer are not obvious, causing easily mistakable
ically essential but challenging. Pulmonary nodule detection is on optimal treatment period. Pulmonary nodules are abnormal
important in early lung cancer detection and diagnosis. False pos- lesions, and they are also signs of suspected lung cancer. Low-dose
itive reduction (FPR) is a significant stage of pulmonary nodule computed tomography (LDCT) imaging is a popular and effective
detection systems. Prior investigations on nodule candidate classi- early lung cancer screening tool. Considering the massive amounts
fication use solitary-nodule approaches, which ignore the relations of data required for diagnosis, manual pulmonary nodule screening
between nodules. In this study, we propose to use Attention-based is laborious and prone to missed diagnosis, especially for small
Deep Multiple Instance Learning, a variation of supervised learn- nodules.
ing to recognize true pulmonary nodules among a large group of In general, pulmonary nodule detection is separated into two
candidates proposed from the detection stage. By treating the mul- problems: candidate detection and then FPR. The goal of FPR stage
tiple nodules from a different patient, critical relational information is to recognize the true pulmonary nodule from the plentiful candi-
between solitary-nodule is extracted and empirically proves the dates received from the first step of pulmonary nodule candidate
benefit of learning the relations between multiple nodules. An at- detection. In pulmonary candidates, there are some similarities
tention layer trained with CNN to replace typical pooling-based between nodules and normal tissues. Therefore, it is demanding
aggregation in multiple instance learning (MIL). Experiments of to predict accurate labels of an individual candidate. Current FPR
lung nodule FPR on the public LUNA16 dataset validate the effec- methods are mainly designed by effective classifiers combined with
tiveness of the proposed method. The proposed method achieved discriminative attributes to reduce the number of candidates. Re-
an accuracy of 99.6%, specificity of 100%, recall of 99.92%, and F1 cently, multiple instance learning (MIL) has been attracting atten-
score of 99.6%. The experimental results reveal that our method can tion as a weakly supervised learning method that can train the
achieve satisfactory performance in FPR. network without making labels on a one-to-one basis.
MIL introduces the concept of "instances" representing individ-
KEYWORDS ual data and "bags" containing all instances belonging to groups.
The network performs the classification task at bag level. In this
Multiple Instance Learning, false positive reduction, pulmonary
case, even if incorrectly predicted instances in the bag will not
nodules, CT images
contribute to performing the final classification. The influence is
ACM Reference Format: low because the final result is performed comprehensively on a bag-
Chi Cuong Nguyen, Giang Son Tran, and Vincent Nguyen. 2023. False Posi- by-bag basis. Recently research has examined the instance-level
tive Reduction of Pulmonary Nodule on CT image using Attention-based integration from a classifier [4, 7, 8]. Among them, [4] proposed
Multiple Instance Learning. In 20th International Conference on Content-
an attention-based deep MIL approach that incorporates two main
based Multimedia Indexing (CBMI 2023), September 20–22, 2023, Orléans,
France. ACM, New York, NY, USA, 4 pages. https://doi.org/10.1145/3617233.
components: attention layer and classification layer. These two net-
3617272 works can concurrently be trained. The attention mechanism helps
the network evaluate information that should be paid attention to
1 INTRODUCTION when the training data contains a large amount of information. By
visualizing the internal parameters of the attention mechanism, it
Lung cancer is the leading cause of cancer death in both males is possible to determine the instances that have been paid attention
and females worldwide. Early lung cancer inspection and diagnosis to and to visualize the basis for the decision. The classification
Permission to make digital or hard copies of all or part of this work for personal or network optimizes learning parameters for the classification task.
classroom use is granted without fee provided that copies are not made or distributed This strategy improves the flexibility and interpretability of the
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
MIL method. By clarifying the information that should be focused
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or on, better performance can be obtained compared to conventional
republish, to post on servers or to redistribute to lists, requires prior specific permission methods.
and/or a fee. Request permissions from permissions@acm.org.
CBMI 2023, September 20–22, 2023, Orléans, France
This paper aims to address the problem of false detection as
© 2023 Copyright held by the owner/author(s). Publication rights licensed to ACM. similarities between nodules and normal tissues in CT images and
ACM ISBN 979-8-4007-0912-8/23/09. . . $15.00 the problem of learning from imbalanced two-class dataset. First,
https://doi.org/10.1145/3617233.3617272
CBMI 2023, September 20–22, 2023, Orléans, France Chi Cuong et al.

we propose to use the attention-based deep MIL approach, which Nodule bags

performs classification at bag level and incorporates the interpreta- Instances


Non-nodule
bags
tion ability with attention mechanism. Second, we consider data
sampling and augmentation to solve the problem of imbalanced
data. We improve the current training algorithm to take into ac- Background
count the skewed distribution of the classes, by giving different Delete

weights to both the majority and minority classes.


The remainder of this paper is organized as follows. Section 2 Figure 2: Generation of instances for nodule and non-nodule
details of our proposed method. Section 3 presents our method’s bags.
experimental setup and results. Finally, we conclude our paper in
Section 4 and present potential directions for future works.
Nodule bags
Attention
2 METHODS CNN ... FC

...
...

...

...
... Mechanism Non-nodule
I1 I2 Ik bags
Output layer
Outline of proposed scheme: FPR in lung nodule detection has Instances Backbone layer h1 h2 hk
MIL Attention
Weighted
layer
Feature vectors
always been demanding in the computer-aided detection system. A feature vectors

high false positive rate reduces the precision of nodule detection.


Because of the objective correlation and largely varied size of nod- Figure 3: Overall structure of the Attention-based deep MIL.
ules from patients, attention-based Deep MIL is more suitable to
address this problem. MIL can learn features of multiple nodules
and non-nodules from different patients; the critical relationship by providing the instance 𝐼𝑘 in the bag into the CNN, and the feature
between multiple instances is extracted at bag level. An outline vector ℎ𝑘 was obtained.
of the proposed method is shown in Fig. 1. Lung CT scans were In the conventional MIL technique, the synthesized vector was
divided into small image patches and stored in bags. Each bag was used for identification. This vector was obtained by calculating the
then labeled as nodule or non-nodule. The final classification will maximum or average value of all components from the feature vec-
be performed by MIL classifier, as nodule bag or non-nodule bag. tor group in each bag. However, the importance of every instance
in a bag could not be accurately identified with the simple math-
Instances
ematics of maximum or average calculation in the conventional
method. Hence, Ilse et al. [4] proposed a MIL technique using an
Nodule
attention mechanism. The attention mechanism allows the CNN
MIL Classifier
to control the attention level on instances most likely labeled as
Non
Nodule nodule and non-nodule. This method can learn critical information
from a large amount of data, consistent with the practical diagnosis
Bag
CT Scan process.
This method uses attention-based CNN for image classification
Figure 1: The overview of our method for False Positive Re- to calculate the attention weight, representing instance importance
duction. level. The weighted feature vector 𝑧 is calculated from a weighted
average of the instances using the following Equations:

Generation of instances: MIL has the ability to learn and clas- 𝐾


∑︁
sify data at bag level. Each CT slice, with a size of 512 × 512 pixels, 𝑧= 𝑎𝑘 h𝑘 (1)
was divided into small patches and used as instance candidates. For 𝑘=1
the input size, we used patches of 64 × 64 pixels as the baseline, where
as the largest nodule can be covered in the image. The CT image
captured the whole lung and therefore contained areas without re- 𝑒𝑥𝑝{w⊤𝑡𝑎𝑛ℎ(Vh⊤
gions of interest (nodules and non-nodule regions). Therefore, the k )}
𝑎𝑘 = (2)
𝐾
image was first divided into regions with nodules/non-nodules and ∑︁

𝑒𝑥𝑝{w 𝑡𝑎𝑛ℎ(Vh⊤
j )}
background regionsGenerally, background regions do not include
𝑗=1
any candidates, with the appearance of other components such as
lung air (black color) and other body tissues (blood vessels, airway where h𝑘 is the feature vector of the 𝑘 𝑡ℎ instance and 𝑎𝑘 is the
passages, ribs, etc.,.). A bag is labeled as non-nodule if all instances attention weight given for each instance in a bag so that the total
in the bag are non-nodule or labeled as nodule if there is at least value of 𝑎𝑘 is normalized to 1 for each bag. Two vectors of w and
one nodule instance. The whole process of instance generation is V are the parameters for calculating the attention weight, which is
described in Fig. 2. determined in training the network. The weighted features z were
Classification method: The structure of the attention-based then assigned to a fully connected (FC) layer with a single artificial
deep MIL [4] used for FPR is shown in Fig. 3. The network com- neuron, and bag classification (nodule and non-nodule) was then
prises a backbone network as CNN architecture, a MIL attention predicted. Here, a sigmoid [2] was used as the activation function.
module, and an output layer. The feature extraction was performed Therefore, attention mechanism weights all instances in the bag to
False Positive Reduction of Pulmonary Nodule on CT image using Attention-based Multiple Instance Learning CBMI 2023, September 20–22, 2023, Orléans, France

create a single vector to represent the bag. This vector was then
identified whether the bag as nodule or non-nodule.
Furthermore, class weighting should be considered to address
the imbalanced data classification problem between nodule and
non-nodule instances. This can be accomplished by giving different
weights to the majority and minority classes. In the training phase,
the weight difference will influence the final classification. However,
weight of each class is usually difficult to choose without prior
knowledge. A natural method of determining weight distribution
is using proportion of classes in training set. The whole purpose
of this method is to penalize the misclassification created by the
minority class by assigning a higher weight value while, at the
same time, reducing the weight of the majority class.

1 𝑛_𝑠𝑎𝑚𝑝𝑙𝑒𝑠
𝑤𝑗 = × (3)
𝑛_𝑠𝑎𝑚𝑝𝑙𝑒𝑠 𝑗 𝑛_𝑐𝑙𝑎𝑠𝑠𝑒𝑠
where 𝑤 𝑗 is weight of each class. 𝑛_𝑠𝑎𝑚𝑝𝑙𝑒𝑠 is total number of
samples in the dataset. 𝑛_𝑐𝑙𝑎𝑠𝑠𝑒𝑠 is total number of unique classes
in the target, here is 2 (nodule and non-nodule classes). 𝑛_𝑠𝑎𝑚𝑝𝑙𝑒𝑠 𝑗
is total number of sample of the respective class.
Figure 4: Example of attention weights for nodule bags.
3 EXPERIMENTS
Image dataset: Our experiment used the public dataset LUng Nod-
ule Analysis 2016 (LUNA16) challenge [9], a widely utilized dataset Experiment Setting: We first build the attention CNN model
for the task of lung nodule detection and FPR. Four experienced with a 64 × 64 input. A convolutional of 64 layers are added to
thoracic radiologists manually annotated the candidate nodules. help the model increase attention on learning instances. Finally,
Each radiologist labeled the nodules as nodules ≥ 3 mm, nodules the network is followed by two fully-connected layers, with 512
< 3 mm, or non-nodules. LUNA16 dataset consists of CT scans from and 256 neurons, respectively. We use an Adam optimizer to train
888 patients with 1,186 nodules, confirmed by at least three of four the CNN with an initial learning rate of 1 × 10 −8 . We exponentially
radiologists. LUNA16 dataset uses the candidate list provided by the decay the learning rate with a ratio of 1 × 10 −2 after every epoch.
competition host, a total of 754,975 candidates as non-nodules. We A lower learning rate can avoid a suboptimal set of weights that
use 1,068 nodules (90%) from the total of 1,186 ground-truth nodules are learned too quickly. The number of training epochs is 100, and
for generating the training and validation set; the remaining 118 the batch size is 1. The training loss is a sparse cross-entropy loss
nodules (10%) are kept for testing data averaged by the number of candidates. In addition, we used early
The dataset provided by the organizers of LUNA16 has about 460 stopping when performance on validation set was not improved.
times more non-nodules than nodules. While applying more train- To confirm the effectiveness of the proposed method, accuracy,
ing samples can improve the training performance, but training specificity, recall, and F1-score are used to evaluate the model pre-
on an imbalanced dataset can lead to overfitting problems. There- diction. We consider the classification ability between nodule and
fore, we implement several sampling and augmentation methods non-nodule of MIL model. We train our model from scratch without
to handle the data skewness problem. We included all the nodule pre-trained weights.
samples in the training data. The training set is further balanced by Results: Table 1 compares the nodule and non-nodule classifica-
up-sampling the nodules by applying the following adjacent slice tion performance of our proposal to other methods. The classifica-
and augmentation methods. In the adjacent slice strategy, most tion by our method, which adopts an attention-deep MIL approach
pulmonary nodules are in spheroidal shapes while radiologists only structure, outperforms all other methods in classification metrics.
annotated CT slices with the largest cross-sectional area. Therefore, The proposed method is the most accurate and received the highest
we also measure two upper and two lower adjacent slices of the recall. A higher recall is especially important in the medical domain
ground-truth nodule image marked by the radiologists and add since false negatives could lead to severe consequences, includ-
them to the training set. This method results in 5,210 lung nodules ing patient fatality. We also notice that the attention mechanism
for the training set. All nodule samples are then augmented to performs better than the plain CNN networks.
increase the total of training samples. Each sample is randomly ro- Finally, we present an exemplary result of the attention mecha-
tated by 20 degrees, horizontally and vertically flipped. In addition, nism of MIL in Fig. 4. In our experiments, each bag consists total of
each nodule image is slightly shifted to a random position. The ran- three images. For each image, the corresponding attention score
dom center shifting method prevents all objects from being located is given by our trained network. The bag is accurately predicted
in the center of the patch. These augmentation methods balance as nodules, and all nodule instances in a bag are correctly high-
the training set. All data sampling and augmentation techniques lighted with higher scores. Hence, the attention mechanism works
were implemented before the training model. as expected.
CBMI 2023, September 20–22, 2023, Orléans, France Chi Cuong et al.

Table 1: Classification performance of our method.

Accuracy Specificity Recall F1-score


Cao et al., 2018 [1] 96.9 98.6 95.1 91.1
El-Regaily et al., 2020 [3] 89.89 90.66 85.26 -
Yuan et al., 2021 [10] 95.6 99.8 81.2 -
Lai et al., 2021 [5] - 98.9 86.4 -
Nguyen et al., 2021 [6] 95.6 97.48 93.84 95.5
Our 99.6 100 99.2 99.6

4 CONCLUSION AND FUTURE WORK Maria Evelina Fantacci, Bram Geurts, et al. 2017. Validation, comparison, and
combination of algorithms for automatic detection of pulmonary nodules in
In this study, we developed a method for classifying nodule and computed tomography images: the LUNA16 challenge. Medical image analysis
non-nodule to reduce the false positive rate in CT images using 42 (2017), 1–13.
[10] Haiying Yuan, Zhongwei Fan, Yanrui Wu, and Junpeng Cheng. 2021. An efficient
attention-based MIL. Experimental show that our attention MIL multi-path 3D convolutional neural network for false-positive reduction of pul-
can outperform the learning ability with supervised learning meth- monary nodule detection. International Journal of Computer Assisted Radiology
ods and highlight the instance that contributed to the decision by and Surgery 16, 12 (2021), 2269–2277.
the attention mechanism. Our method automatically classifies nod-
ule candidates in CT images more efficiently and accurately than
traditional supervised learning methods. Our sampling and class
weighting can address the problem of imbalanced learning.
We used a binary classification problem to evaluate our proposed
approach. However, extending the current study to a multiple-class
classification task as nodule malignancy prediction through an
attention-based MIL network is an interesting direction for fur-
ther research. In addition, by taking advantage of three-dimension
spatial information of nodules in CT images, the three-dimension
attention-based MIL method might be one potential shortcoming
to be overcome in future research.

ACKNOWLEDGMENTS
This work was supported by the French Embassy in Vietnam under
the grant No. 16/2022/SCAC.

REFERENCES
[1] Guitao Cao, Tiantian Huang, Kai Hou, Wenming Cao, Peng Liu, and Jiawei Zhang.
2018. 3D convolutional neural networks fusion model for lung nodule detection
onclinical CT scans. In 2018 IEEE International Conference on Bioinformatics and
Biomedicine (BIBM). IEEE, 973–978.
[2] George Cybenko. 1989. Approximation by superpositions of a sigmoidal function.
Mathematics of control, signals and systems 2, 4 (1989), 303–314.
[3] Salsabil Amin El-Regaily, Mohammed Abdel Megeed Salem, Mohamed Has-
san Abdel Aziz, and Mohamed Ismail Roushdy. 2020. Multi-view Convolutional
Neural Network for lung nodule false positive reduction. Expert systems with
applications 162 (2020), 113017.
[4] Maximilian Ilse, Jakub Tomczak, and Max Welling. 2018. Attention-based deep
multiple instance learning. In International conference on machine learning. PMLR,
2127–2136.
[5] Khai Dinh Lai, Thuy Thanh Nguyen, and Thai Hoang Le. 2021. Detection of lung
nodules on ct images based on the convolutional neural network with attention
mechanism. Annals of Emerging Technologies in Computing (AETiC) 5, 2 (2021),
78–89.
[6] Chi Cuong Nguyen, Giang Son Tran, Jean-Christophe Burie, Thi Phuong Nghiem,
et al. 2021. Pulmonary nodule detection based on faster R-CNN with adaptive
anchor box. Ieee Access 9 (2021), 154740–154751.
[7] Jan Ramon and Luc De Raedt. 2000. Multi instance neural networks. In Proceedings
of the ICML-2000 workshop on attribute-value and relational learning. 53–60.
[8] Vikas C Raykar, Balaji Krishnapuram, Jinbo Bi, Murat Dundar, and R Bharat
Rao. 2008. Bayesian multiple instance learning: automatic feature selection and
inductive transfer. In Proceedings of the 25th international conference on Machine
learning. 808–815.
[9] Arnaud Arindra Adiyoso Setio, Alberto Traverso, Thomas De Bel, Moira SN
Berens, Cas Van Den Bogaard, Piergiorgio Cerello, Hao Chen, Qi Dou,

You might also like