Journal of Ambient Intelligence and Humanized Computing

https://doi.org/10.1007/s12652-022-03835-8

ORIGINAL RESEARCH

MSDNet: a deep neural ensemble model for abnormality detection

and classification of plain radiographs
K. Karthik1   · S. Sowmya Kamath1 

Received: 30 June 2021 / Accepted: 28 March 2022

© The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2022

Abstract
Modern medical diagnostic techniques facilitate accurate diagnosis and treatment recommendations in healthcare. Such diag-
nostics procedures are performed daily in large numbers, thus, the clinical interpretation workload of radiologists is very high.
Identification of abnormalities is a predominantly manual task that is performed by radiologists before the medical scans are
available to the patient’s referring doctor for further recommendations. On the other hand, for a radiologist to delineate the
imaging study’s findings/observations as a textual report is also a tedious task. Automated methods for radiographic image
examination for identifying abnormalities and generating reliable radiology report are thus a fundamental requirement in
clinical workflow management applications. In this work, we present an automated approach for abnormality classification,
localization and diagnostic report retrieval for identified abnormalities. We propose MSDNet, an ensemble of Convolutional
Neural models for abnormality classification, which combines the features of multiple CNN models to enhance abnormality
classification performance. The proposed model also is designed to localize and visualize the detected abnormality on the
radiograph image, based on an abnormal region detection algorithm to further optimize the diagnosis quality. Furthermore,
the extracted features generated by MSDNet are used to automatically generate the diagnosis text report using an automatic
content-based report retrieval algorithm. The upper extremity musculo-skeletal images from the MURA dataset and chest
X-ray images from Indiana dataset were used for the experimental evaluation of the proposed approach. The proposed model
achieved promising results, with an accuracy of 82.69%, showing its significant impact on alleviating radiologists’ cognitive
load, thus improving the overall efficiency.

Keywords  Abnormality detection · Health informatics · Medical image classification · Automated report generation

1 Introduction and pneumothorax. As per statistics published by the United

X-ray imaging serves as a valuable tool for a wide variety
of medical examinations and diagnostic procedures. One
major application is the detection of bone fractures, identi-
fication of types of injuries and existence of foreign objects
in the concerned area etc. In chest disease diagnosis, X-rays
play a major part in diagnosing diseases like pneumonia,
cardiomegaly, pulmonary edema, opacity, pleural effusion,
* K. Karthik
2karthik.bhat@gmail.com
S. Sowmya Kamath
sowmyakamath@nitk.edu.in
1
Healthcare Analytics and Language Engineering (HALE)
Lab, Department of Information Technology, National
Institute of Technology Karnataka, Surathkal, Mangaluru,
Karnataka 575025, India
States Bone and Joint Initiative (BMUS 2014), more than 1.7
billion people worldwide are affected by musculo-skeletal
conditions. It is also reported that the occurrence of chronic
musculo-skeletal conditions in adults is significantly more
than that of coronary and heart conditions, and double that
of chronic respiratory conditions, which are ranked at the
second and third position as the most common health condi-
tions in the world. Musculoskeletal conditions cause severe
long-term pain and disability (Woolf and Pfleger 2003), and
it has been reported that an average of 30 million people
visit hospitals every year. This number has steadily increased
year on year, thus causing significant surge in the volume of
radiology scans ordered by referring doctors for enabling
diagnosis. At present, identification of abnormalities is a
highly manual task that is performed by radiologists. The
medical scans are forwarded to the patient’s referring doctor
for further recommendations. A radiologist’s routine work

13
Vol.:(0123456789)

thus requires handling a large amount of data, and the possi-
bility exists that the intensive labor may lead to misdiagnosis
(Ying et al. 2016). A study carried out to measure radiolo-
gist exhaustion during the process of interpreting musculo-
skeletal radiographs (Krupinski et al. 2010), reported that
a significant decrease in their overall efficiency during the
process of fracture detection from diagnostic scans, at the
end of the work day when compared to the beginning.
Musculo-skeletal Disorders (MSDs) are often debilitat-
ing and affect a person’s daily activities to a large extent.
Hairline fractures, carpal tunnel syndrome, ligament tear are
some examples of such disorders, the potential causes of
which could be work-related accidents, workplace fatigue,
sport injuries etc. The daily routines of busy professionals
like those who do repetitive tasks on computers (banking,
software professionals etc), have seen increasing occur-
rence of neck and lower back pain. Researchers that studied
occupation-related cases of such professionals observed that
physical and individual factors affected the health of the
worker’s musculo-skeletal system over time (Silvian et al.
2011; Wærsted et al. 2010). Playing Related Musculo-skel-
etal Disorder (PRMD) is a common ailment affecting career
musicians, it has been reported that string quarter players
experience more MSDs than other instrumentalists (Baadjou
et al. 2016). Early diagnosis can help and emerging diagnos-
tic procedures like physiotherapy, shock wave therapy etc
have been used to treat any mild to medium symptoms of
pain caused by MSDs. Primarily, such treatments can help
give some relief in disorders like tendonitis and osteoarthri-
tis (which cause pain, swelling in the elbow, wrist, finger,
thigh, and other parts of the body, due to overuse) and bone
fractures (Ioppolo et al. 2014).
Early detection and effective diagnosis for assessing any
abnormalities in scans, is thus crucial for further treatment.
Currently, such abnormality detection is predominantly a
manual task performed by expert radiologists or orthopae-
dists, making it a time-consuming process demanding sig-
nificant time and effort from trained healthcare profession-
als. Given the large volume of diagnostic scans undergone
by patients each day in hospitals, this is a significant burden
on available healthcare resources, often delaying treatment
(Dalia et al. 2021). In view of this, automated algorithms for
classification and detection of anomalies is a crucial require-
ment and can revolutionize the ways in which healthcare can
be delivered at the point-of-care.
Abnormalities that exist in medical images can be char-
acterized and detected in various ways. A traditional and
common categorization is using a shape model. Certain
abnormalities have unique geometric shapes, and their sizes
are helpful for detection of disease staging and prognosis.
Considering an example, some tumor regions are round in
shape and tend to grow in size. In such cases, detection of
abnormalities can be considered as a task with particular
13
K. Karthik, S. Sowmya Kamath
pattern in the medical image. These patterns can be from
a collection of geometric shapes, like corners, edges, lines,
circles, contours, etc. These shapes are detected by means of
a specific method or technique, for example, fracture lines
can be detected by various edge detection techniques. How-
ever, these can be obscured in some scan view orientations,
affecting the detection performance Karthik and Kamath
(2021b). Further, the performance of detecting the abnor-
malities can be improved by integrating prior knowledge
of abnormalities, such as the location, size or orientation.
With the advancement of applications of artificial intelli-
gence (AI) based models for healthcare, research in Machine
learning (ML) and deep learning (DL) have shown great
promise in clinical tasks like disease diagnosis (Mukesh
et al. 2021; Katara et al. 2021), medical image classification
(Faes et al. 2019; Kumar et al. 2016; García-Floriano et al.
2019) and image retrieval (Mandikal et al. 2019; Soundal-
gekar et al. 2018; Karthik and Kamath 2021a) among oth-
ers. The performance of these systems in clinical settings
can revolutionise the way healthcare services are delivered,
especially in a labor-intensive field like radiography, where,
the radiologist is expected to manually check each scan and
write a list of observations, for enabling diagnosis by the
referring doctors.
Image processing and Computer Vision (CV) based
techniques have been applied for designing applications
for surgical and imaging interventions. Such systems are
extend clinical decision making capabilities to the healthcare
professionals, by automating certain tasks related to diagno-
sis, or by forecasting the severity of several abnormalities
(Nedumkunnel et al. 2021; Mayya et al. 2021). Incorporat-
ing AI in these systems to support learning behaviour so
that systems can detect abnormalities at the earliest disease
onset in a wide variety of diagnostic media are of critical
importance. The radiologist can utilize these insights for
enabling and optimizing the quality of diagnosis (Katara
et al. 2021). The marked region can help the physicians to
focus on early and effective treatment recommendations.
Further, automated retrieval of radiological diagnosis reports
will minimize the manual work involved in reporting obser-
vations from the radiological images, while also alleviat-
ing the cognitive burden of the radiologists due to the huge
load of cases that they typically handle each day. The major
contributions of the research work presented in this article
are listed below:
1. Design of ensemble deep neural model for classifying
abnormal radiography images.
2. Design of a boundary detection algorithm for identifying
the region of interest for facilitating anomaly detection.
3. A content-based medical report retrieval approach for
automatically generating diagnostic text reports for a
MSDNet: a deep neural ensemble model for abnormality detection and classification of plain…
given image based on the anomalies identified by the
proposed models.
The rest of this paper is organized as follows: In Sect. 2, we
present a detailed discussion on available recent work in the
area of interest. Section 3 details the proposed methodology
for abnormality classification, localization and text retrieval
for different radiograph classes. The details of experiments
conducted for benchmarking the proposed approaches are
presented in Sect. 4, followed by conclusion and future work.
2 Related work
Recent advancements in machine learning and deep learn-
ing architectures have emphasized the potential improve-
ments in classification tasks, with major impact in the field
of diagnostic image analysis and management. The avail-
ability of large diagnostic image datasets have also helped
in the development of deep learning models for different
tasks such as image classification, recognition, segmentation
and so on. Efforts are in progress for developing large, open
and standard repositories for facilitating medical image man-
agement. The MURA dataset made available by Rajpurkar
et al. (2017a) consists of an extensive collection of musculo-
skeletal studies of upper extremities like hand, wrist, elbow,
finger, forearm, humerus and shoulder radiography classes.
The Indiana university dataset (Rajpurkar et al. 2017b) pro-
vides chest X-rays with textual reports reporting observed
abnormal conditions.
Automatically deciding if a radiography study is normal
or abnormal is a task that can significantly reduce the bur-
den of radiologists who handle a large volume and variety
of radiological images each day. Rajpurkar et al. (2017a)
developed a deep 169-layer convolutional neural network
for predicting the abnormality of musculoskeletal studies.
The model predicts a study as abnormal if the prediction rate
was found to be greater than 0.5. Here, the variable-sized
images were scaled to 320×320 and the network weights
were set according to the pre-trained ImageNet model with a
learning rate of 0.0001 and minibatch size of 8. Their results
performed best for wrist and humerus classes using Cohen’s
kappa statistic with a 95% confidence interval compared
with the three best radiologists.
Saif et al. (2019) developed a model based on the Capsnet
architecture called a capsule network for identification of
abnormalities present in musculoskeletal radiographs. By
varying the image size to different pixel sizes, the perfor-
mance of the network was observed at each cases. Also the
proposed capsnet model achieved an improved kappa score
of 10% when compared to the 169 layered Densenet model,
while using 50% less training data. Banga and Waiganjo
(2019) proposed an ensembled 200 model for abnormality
detection using musculoskeletal radiographs. Their model’s
performance was not satisfactory as its Cohen Kappa score
was lesser than that of the DenseNet169 model (Rajpurkar
et al. 2017a), but in terms of F1 score, it outperformed the
DenseNet model. Solovyova and Solovyov (2020) proposed
a 169 layered Deep CNN model for abnormal X-ray image
classification. To increase the quality of the initial pre-pro-
cessing for identifying the region of interest, they used dif-
ferent augmentation techniques. Their model performed well
on bone abnormality classification, however, it is limited to
the binary classification problem, i.e., whether an anomaly
exists or not.
An ensemble model built on Inception V3, ResNet,
and Xception was developed by Kitamura et al. (2019).
The model’s performance was measured considering the
multiple views of the ankle radiographs, to determine the
effect of fracture detection. However, the model was trained
and tested on a small dataset consisting of 596 normal
and abnormal ankle cases. Chung et al. (2018) proposed
a deep CNN after augmenting the training dataset to clas-
sify proximal humerus fractures. They claimed that their
model showed superior performance when compared with
the human groups of general physicians and orthopedists.
However, their study neglected to address the clinical task
of fracture region detection, which is of primary importance.
Gale et al. (2017) proposed a hip fracture detection approach
based on deep neural networks on a small annotated data-
set which includes 3,354 images, with 348 fractures, and
reported good results.
Yahalomi et al. (2019) proposed a Faster R-CNN based
model for identification of fractures in anteroposterior hand
X-ray images. Their model was reported to be more accurate
in detecting the fracture regions than by the radiologist’s
findings, however, the model used a small set of diagnosed
X-ray images and was pre-trained on the Imagenet dataset.
Kim and MacKinnon (2018) used Inception v3 as a trans-
fer learning approach for deep CNNs for classifying wrist
radiographs into fractured/no-fracture classes. However,
they used a very small dataset and thus their approach is not
scalable. Krogue et al. (2020) proposed a DenseNet deep
learning model for identification and classification of hip
fractures. The model considered a single class image for
prediction, unlike a human interpreter, who may look at sev-
eral views for abnormality classification. Cheng et al. (2019)
used a deep CNN model to detect and visualize hip fractures.
Further, their model used Grad-CAM models to check the
model’s prediction by highlighting the fracture region.
Tataru et al. (2017) used three different kinds of CNN
models like GoogLeNet, InceptionV3 and ResNet, for
classifying chest X-rays as either normal, abnormal, or
extremely high-risk. The original image sizes were down-
sampled to 512*512 and 224×224 from 3000×3000 pixels
before feeding them to the neural model. Among the three,
13

GoogleNet performed the best for the chest X-ray image
classification task. Adding to this, detecting the abnor-
mal region or including the text reports already available
for chest X-rays in the generation of the reports could be
added in their methodology pipeline. Aowal et al. (2017)
employed DCNN based models for abnormality detection
in chest X-rays. Their approach showed consistent detection
results even when the number of available training examples
were low, and when random & multiple train-test data splits
were used. However, the DCNN did not perform well across
all abnormalities. Their study considered only frontal chest
X-ray images and text reports were also not incorporated.
Harzig et al. (2019) proposed a dual-word LSTM sentence
generation model separately for abnormal and normal chest
X-ray images. They reported that dual-word LSTMs helped
increase the number of distinct sentences, but, failed to iden-
tify abnormal regions in the image.
Table 1 summarizes the major research works related
to abnormality classification, localization of the abnormal
region and text report summary for a given image under
test. From the extensive study of existing works under-
taken, we observed that the problem of classification and
identification of abnormalities still evince significant scope
for further research. Several challenges exist that need to
be addressed, introduced due to the multimodal nature of
diagnostic images, i.e., the complexity of the image and the
extensive variations across different imaging modalities.
After an image is classified as abnormal, identifying the
fracture region or detecting the area of abnormality has not
received much attention. It was also observed that, diagnos-
tic text report generation based on the observations obtained
from the underlying neural models is under-explored. In the
overall process of radiography image examination, the inclu-
sion of the identified findings in the form of an automatically
generated textual report would be an added advantage for
physicians.
In view of these observations, our work encompasses
developing an ensemble deep neural network model for
automated abnormality detection and classification. The
abnormalities present in the images are identified using our
developed abnormal region detection algorithm. Further, the
features generated by the ensemble neural model are used for
the auto retrieval of radiological text reports, thus reducing
the radiologist’s workload and making it more manageable.
3 Proposed approach
The proposed approach for abnormality classification,
abnormality localization and automated diagnostic report
retrieval we used two publicly available standard datasets.
The first one, the MURA dataset (Rajpurkar et al. 2017a)
provides musculoskeletal radiograph images of seven upper
13
K. Karthik, S. Sowmya Kamath
extremity classes like Hand, Forearm, Wrist, Finger, Shoul-
der, Humerus and Elbow. The second dataset, Indiana Uni-
versity dataset consists of chest X-ray images, along with
indications, findings, and impressions in a textual form
for each image. These two datasets were together used for
addressing the three different clinical tasks, each involving
data belonging to two data modalities, X-ray images and
clinical text reports, thus resulting in multimodal data-
sets. Due to this, the proposed ensemble models are more
robust, capable of dealing with varied types of abnormali-
ties in underlying radiographic images. The three clinical
tasks undertaken form a significant part of a typical clinical
workflow that is managed on a daily basis by a radiologist
in hospital scenarios
1. Classifying a given diagnostic image as either normal or
abnormal.
2. Abnormal region detection for localize and visualizing
identified abnormal areas.
3. Automated diagnostic text report generation.
The proposed methodology is illustrated in Fig.  1. The
images are first pre-processed, after which they are fed into
the proposed neural network model for abnormality classi-
fication. For preprocessing, the images present in the data-
sets are of different sizes. The height and the width of each
image are not uniform, hence these are resized to a standard
size of 256 × 256 for AlexNet and 224 × 224 for ResNet
to train our models efficiently. The input size was selected
based on the input layer of the neural network to preserve
the information as much as possible and simultaneously to
reduce the overall computational cost during training the
model. Any abnormal regions are identified by the proposed
abnormal region detection algorithm by marking a bounding
box around the identified region. Next, the task of automatic
retrieval of diagnosis report for the given image under test,
is performed. The methodology adopted for addressing each
of these tasks are discussed in more detail in the subsequent
sections. The proposed MSDNet ensemble model is the core
of the pipeline that is used for the three clinical tasks, which
is trained to extract features and image index values that will
facilitate the classification, identification and localization of
anomalies, and also automated diagnostic text report genera-
tion for a given input image.
3.1 Abnormality classification
The architecture of the proposed MSDNet model is depicted
in Fig. 2. The proposed network is built as an ensemble of
the AlexNet (Krizhevsky et al. 2012) and ResNet18 (He et al.
2016) architectures for initial classification of the category of
the image, i.e. abnormal or normal. The global features are
obtained from the first model by directly feeding the original
 Table 1  Summary of key works on abnormality detection and report generation
Author & Year
Rajpurkar et al. (2017a)
Tataru et al. (2017)
Aowal et al. (2017)
Kim and MacKinnon (2018)
Chung et al. (2018)
Saif et al. (2019)
Banga and Waiganjo (2019)
Chada (2019)
Yahalomi et al. (2019)
Kitamura et al. (2019)
Cheng et al. (2019)
Krogue et al. (2020)
Harzig et al. (2019)
Solovyova and Solovyov (2020) Pre-processing and deep 169 layer CNN based
13
Methodology Merits Observed limitations
Deep CNN model with 169 layers Abnormality prediction if threshold >0.5, binary Only Cohen’s kappa statistic used for evaluation
classification
GoogLeNet,InceptionV3 and ResNet Visualization results demonstrate that macroscopic Abnormal region detection is not performed
features are learned effectively by the model
Different CNN based models Showed good consistent detection results when the Models did not perform well across all abnormali-
number of training examples is low ties. The study considered only frontal chest X-ray
images
Inception v3 Transfer learning approach, pre-trained on non- Initial dataset size 1,389 wrist images was increased
medical images, applied to the problem of to 11,112 using data augmentation techniques
fracture detection in plain radiographs
Deep CNN CNN showed a high performance and when com- Small dataset of 1,891 humerus images was used in
pared human groups, the CNN showed superior the study
performance
Capsule network architecture Performance of the network features vanquished No abnormality region detection
the limitations of CNN
Ensembled200 model The model outperformed DenseNet model Cohen Kappa’s score was not computed. No
bounding-box or abnormality region detection
DenseNet-169, DenseNet-201 and Inception- All three models showed good performance on Only two classes were used for the study, other
ResNetV2 models humerus and finger class of MURA radiograph classes not evaluated
images
Faster R-CNN Achieved an accuracy of 96% in identifying frac- A small dataset including hand X-ray images (single
MSDNet: a deep neural ensemble model for abnormality detection and classification of plain…
tures for augmented antero-posterior hand X-ray class) was used and pre-trained on the Imagenet
images dataset
Ensemble Models (Inception V3, Resnet, Xception Models used 3 views for each case and achieved an Only one single class which included 298 normal
CNNs) accuracy of 81% and 298 fractured ankle studies were used
DCNN and Grad-CAM Localizing fracture lesions Only four classes, Ankles, elbows, feet, and wrists,
were used in the study
DenseNet and bounding box Classification and identification of the model for Limited data size which included only 1849 images
hip fractures achieved with at least expert-level of a single class in its study
accuracy
Dual word LSTM sentence generation model Results proved that the dual word LSTM helped to Did not capture the findings or identification of
increase the number of distinct sentences abnormal regions in the image
The model showed a good results in bone abnor- focused only on the binary classification problem
neural model malities detection

Fig. 1  Proposed methodology for abnormality classification, detection and report formation
Fig. 2  Architecture of the proposed MSDNet model
input image, while ResNet18 is used to generate the local fea-
tures, for a pre-processed image fed into the neural network.
A pre-processed image is obtained by finding the maximum
connected components of a binary image. For this, firstly, the
convex-hull of the image is generated and the obtained binary
image is superimposed on the original image to obtain the
maximum connected component. Global and local features
are then combined to form a composite feature set. These con-
catenated features are then fed into the fully-connected layer
for final abnormality classification. For this purpose, the final
three layers were adjusted as per the requirements of the clas-
sification task. Here, the last three layers were replaced with
a fully connected layer, a softmax layer, and a classification
output layer. The final fully connected layer was redesigned
according to the number of classes in the new data for clas-
sification outputs. The images are fed into the network for gen-
erating the probability of occurrence of any abnormalities in
each image. If the probability value is equal to or higher than
0.5, then the image is classified as an abnormal study. Here, the
task is to classify the given image into the category to which
it belongs. Hence, it is a single-label categorical classification
(i.e., softmax activation), and the standard weighted categori-
cal cross-entropy loss is given by:
13
K. Karthik, S. Sowmya Kamath
K M
1 ∑∑
Jwcce = − w × ykm × log(h𝜃 (xm , k)) (1)
M k=1 m=1 k
where, M gives the number of training examples; K gives
the number of classes; wk is the weight for class k; ykm is the
target label for training example m for class k; xm is the input
for training example m and h𝜃 represents a model with neural
network weights 𝜃.
For the training process, we experimented with different
batch sizes like 8, 16, 32, 64 and finally the learning rate
was set to 0.0001, with batch size of 16 and the number of
epochs as 10. Stochastic Gradient Descent with Momentum
(SGDM) was used as the solver optimizer as it can switch
back and forth to reach the optimum path. Hence, momen-
tum parameters were added to reduce the switching problem
(McHugh 2012). The values of SGDM are calculated and
updated as per Eq (2), where, 𝓁 is the iteration number, 𝛼 > 0
is the learning rate, 𝜃 is the parameter vector, E(𝜃 ) is the loss
function and 𝛾 determines the contribution of the previous
gradient step to current iteration.
𝜃𝓁+1 = 𝜃𝓁 − 𝛼∇E(𝜃𝓁 ) + 𝛾(𝜃𝓁 − 𝜃𝓁−1 ) (2)
MSDNet: a deep neural ensemble model for abnormality detection and classification of plain…

To avoid considering the same data segments at every epoch,

a shuffling parameter value was set before each training
Algorithm 1 Abnormal Region Detection
epoch. i.e., if the mini-batch size cannot uniformly distribute
the training samples, then trainNetwork discards the training Input: An Abnormal Image
data that does not fit into the final complete mini-batch of Output: Boundary region on Abnormal
each epoch. To learn faster in the new layers, we increased Image
the WeightLearnRateFactor and BiasLearnRateFactor val- 1: for i = 1 to length(TestImage) do
ues of the fully connected layer. This was heuristically set 2: IT ← TestImage  Read the test image
to 20, the learning rate is determined by multiplying this 3: Compute the histogram of the image.
factor by the global learning rate, for deriving the biases in 4: Find total histogram values > 1000.
the fully connected layer. The cross-entropy loss function is 5: binaryImage ← grayImage < x-axis value
used as an encoded output. For a single image this loss is with last peak from histogram  Binarization
computed as per Eq. (3), where, M is the number of classes 6: binaryImage ← bwareafilt(binaryImage,
and ŷc is the model’s prediction for that class (i.e. the output 2)  Extract only the two largest blobs
of the softmax for class c). y is a (2× 1) vector of one’s and 7: labeledImage ← bwlabel(binaryImage) 
zero’s, yc is either 1 or 0. Finally, the predicted class label of Label Connected Components
an input X-ray image is obtained at the output layer. 8: binaryImage2 = labeledImage ← 0
M 9: binaryImage2 ← imfill(binaryImage2,
’holes’)  Fill holes
∑
Cross − entropy loss = (yc ⋅ logŷc ) (3)
c=1 10: TestImage(∼ binaryImage2) ← 0  zero
out the other parts of the image
3.2 Abnormal region detection 11: Mask ← grayImage > call  Get a new
binary image
After an image is classified as normal or abnormal, an auto- 12: Mask ← imfill(Mask, ’holes’)  Fill holes
mated analysis of the type of abnormality present in the 13: Mask ← bwareafilt(Mask, 1)  largest
image is of critical importance. The objective is to identify blob selected
potential abnormality findings like hardware artifacts and the 14: Mask ← bwconvhull(Mask)  Take
existence of fractures in the scan image. Boundary detection convex hull
in abnormal images is one of the crucial steps while gener- 15: Get the Centroid points.
ating X-ray scan reports. Currently the abnormal regions 16: Mark bounding box & centroid points for
are manually marked by expert clinicians/physicians. The abnormal regions
developed algorithm can make a change over in the Com- 17: Output abnormalities.
puter Aided Diagnosis medical system, where the boundary 18: end for
will be marked by the system itself if it finds any abnormali-
ties present in the image. In our work, a boundary detection 3.3 Automatic diagnostic report generation
algorithm is incorporated to detect the abnormalities present
in the image. Algorithm 1 illustrates the process of identify- For this task, we attempt to use the modeled features
ing the abnormal regions in the radiograph images. For each extracted from the X-ray images and the expert-written
image, the histogram of the abnormal image is plotted and diagnosis reports for capturing the findings and impres-
the last peak value along the x-axis of the histogram value is sions as a text report for the identified abnormal chest
used to find the two largest objects/blobs on the image. The X-rays taken from the Indiana University dataset.
binarized image with less than the bin location is then used The convolutional layers that build up each of the two
to determine the largest blobs/regions in the image. Next, adapted models based on the architectures of AlexNet and
the connected components of these two blobs are obtained, ResNet-18 were trained to extract disease-specific features
after which the centroid points are found from the convex from the X-ray images. Using these extracted image fea-
hull masked image. Using the centroid points, a bounding tures and the image ID’s, we tried to incorporate the find-
box is marked around the abnormal region of the image. ings and impressions of the image readings as a natural

13
 K. Karthik, S. Sowmya Kamath

Fig. 3  Automatic report generation process for chest X-ray images

language text report for a test chest X-ray image. During

experiments, it is observed that the ensembled high-level
Algorithm 2 Automatic Report Retrieval Algo-
image feature set contributed more to achieve the best
rithm
performance for the inclusion of text reports compared
Input: A Sequence of Images and its corre-
to the individual features from ResNet and AlexNet. Fig-
sponding Report.
ure 3 illustrates the process of automatic diagnostic report
Output: Image and its corresponding
retrieval.
Report.
Algorithm 2 shows the report retrieval process after the
image classification process. During feature extraction, each 1: For each test feature set

image in the training and test sets is processed for generat- 2: ID’s ← Find the nearest distance in train

ing a feature vector. When a test image feature vector is feature set.
given as a query, the pairwise distance measure is used to 3: Read text report file.  Image’s Report.
compute distances to obtain the matching feature vector with 4: Text ← IID , Indication, Findings & Impres-
the smallest distance from the images in the training set. For sion.
each observation in Y (Test image features), the pairwise 5: for i = 1 to length(TestImage) do
distance method finds the smallest distances by computing 6: I ← TestImage  Read the test image
and comparing the distance values to all the observations in 7: Il ← TestImageLabel  Abnormality label
X (Training image features). It was observed that Cosine and of test image.
Standard Euclidean achieved the smallest distance measure 8: for j = 1 to length(ID) do
with an overall accuracy of 78.03% and 77.26%. Hence, for 9: MI ← TrainImage(ID[i,j])  nearest
this report generation experiment, we used Cosine as the dis- matched train image.
tance measure. When a test image is fed into the model, the 10: Lbl ← TrainImageLabel(ID[i,j]) 
Cosine distance is computed to find the training set’s nearest Abnormality label of matched image.
match. Its equivalent index number is the closest reference, 11: Index ← ID[i, j]  Index number of
using which the image and its description (Indication, Find- matched image.
ings & Impression) are retrieved and displayed. 12: Ind = Text(Index,a)
13: Find = Text(Index,b)
14: Imp = Text(Index,c)  Store
Indication, Findings & Impression of the rel-
evant Index row., where a, b & c are Indica-
tion, Findings & Impression column number
of Report file.
15: end for
16: Display I, Il , MI , Lbl, Ind, Find and Imp.
17: end for

13
MSDNet: a deep neural ensemble model for abnormality detection and classification of plain…

4 Experimental evaluation hardware artifacts, degenerative joint diseases and other

For the experimental evaluation of the proposed model two
publicly available datasets were used - the MURA and Indi-
ana Datasets. The MURA dataset (Rajpurkar et al. 2017a)
is provided by the Stanford Group, and consists of mus-
culoskeletal radiography images of seven upper extremity
classes. The scan images are represented as per the Picture
Archive and Communication Systems (PACS) specifica-
tions. The MURA dataset consists of 14,863 studies taken
from 12,173 patients, consisting of 40,561 radiography
images in total. The dataset is split into training (36,808
images, 13,457 studies from 11,184 patients), testing (3,197
images, 1,199 studies from 783 patients) sets. There is no
overlap in patients between any of the sets and each study
has been labeled manually as either normal or abnormal,
by Stanford Hospital’s board-certified radiologists. Abnor-
mal images in the dataset contain anomalies like fractures,
miscellaneous abnormalities, including lesions and subluxa-
tions. A sample of normal and abnormal images is shown in
Fig. 4. The Indiana dataset (Demner-Fushman et al. 2016)
includes 7,470 chest X-Ray images having both frontal and
lateral images with annotations, which consist of indications,
findings, & impressions in a textual form. We use this data-
set for the clinical task of abnormality classification and for
retrieving the reports for a given test image. Hence, only the
frontal chest X-ray images (comprising of 4,000 images) to
extract the image’s relevant features at the training phase. A
sample of this is shown in Table 2.
First, the predicted class labels are obtained from the deep
ensemble model. It is observed that, for the seven classes of
the MURA dataset, a total of 3,197 images - 1,667 images
were normal and 1,530 images contain abnormal findings
(fractures, hardware artifacts and joint diseases). In the Indi-
ana dataset, abnormal/ disease annotations like cardiomegaly,
opacity, pleural effusion, pneumothorax, pulmonary edema,

Fig. 4  Sample images of MURA dataset in hand, forearm, wrist, finger, shoulder, humerus and elbow classes (Upper row—abnormal images;
Lower row—normal images)

Table 2  A sample of Indiana dataset chest X-ray images and its description

Image Indication Findings Impression

Preoperative renal transplant The lungs and pleural spaces show no acute abnormal- No acute pulmonary findings. Mild cardiomegaly
Chest and midback pain
Shortness of breath
ity. Stable left upper lobe calcified granuloma. Heart
size is mildly enlarged, pulmonary vascularity within
normal limits. Mild tortuosity of the descending
thoracic aorta
Stable cardiomediastinal silhouette with tortuous Stable exam with no acute abnormality seen
thoracic aorta. No pneumothorax, pleural effusion or
suspicious focal air space opacity. Stable right lung
base scarring
The cardiac contours are normal. The lungs are Emphysema without superimposed pneumonia.
hyperinflated with flattening of the diaphragms and
tapering of the distal pulmonary vasculature. There
is no focal consolidation. Thoracic spondylosis. Mild
dextroscoliosis of the spine. Prior anterior cervical
fusion

13
 K. Karthik, S. Sowmya Kamath

Table 3  Classification accuracy w.r.t AlexNet, ResNet18 and pro- Table 4  Classification performance w.r.t different classes for the pro-
posed MSDNet model posed MSDNet model
Classes AlexNet ResNet18 MSDNet Classes Accuracy Sensitivity Specificity Kappa

Elbow 0.7867 0.8218 0.8317 Elbow 0.8317 0.8383 0.7004 0.736

Finger 0.7113 0.8059 0.7994 Finger 0.7994 0.7897 0.7263 0.671
Forearm 0.7450 0.8112 0.8394 Forearm 0.8394 0.8533 0.6894 0.792
Hand 0.7082 0.7603 0.7832 Hand 0.7832 0.8413 0.6732 0.754
Humerus 0.7985 0.8549 0.8586 Humerus 0.8586 0.7905 0.8086 0.676
Shoulder 0.6956 0.7937 0.7935 Shoulder 0.7935 0.6877 0.7635 0.731
Wrist 0.8067 0.8469 0.8447 Wrist 0.8447 0.8874 0.7458 0.855
Chest 0.8221 0.8571 0.8827 Chest 0.8827 0.8911 0.7231 0.758
Overall 0.7625 0.8218 0.8269 Overall 0.8269 0.8179 0.7662 0.746

fed throught the subsequent layers, to finally produce the

and shortness of breath were identified in the study. Of these
sets, 81.79% of images were correctly classified as normal
and 76.62% of images were predicted to be abnormal. The
proposed approach’s performance is measured using standard
evaluation metrics like accuracy, sensitivity, specificity, and
kappa statistics. Sensitivity is a measure of the true positive
rate or probability of detection, i.e., it indicates the percent-
age of medical scans correctly identified as abnormal. Speci-
ficity (True Negative Rate) gives the percentage of normal
medical scans correctly classified as normal.
Cohen’s kappa statistics (McHugh 2012) measures the
precision or inter-observer agreement, and is especially use-
ful when similar score is assigned to the same data items.
Hence, the outcome is to predict whether the data sample
under test is normal or abnormal. The importance of this
metric lies in the correct representation of the data meas-
ured. Its range is -1 to +1; a value of 1 indicates a “perfect
agreement” and a value less than 1 shows “less than a per-
fect agreement”. In some rare situations, the Kappa value
can also be negative, signifying that the agreement score
is much lower than expected. It is computed as per Eq. (4),
where Pr(a) is the actual observed agreement, and Pr(c) is
the chance agreement.
Pr(a) − Pr(c)
Kappa = (4)
1 − Pr(c)
The results of the experimental evaluations conducted
with each neural model selected for the comparison are tabu-
lated in Tables 3 and 4. From these results, it was observed
that ResNet outperformance AlexNet, which can be attrib-
uted to its 71 layer deep architecture, in contrast to AlexNet’s
25 layer architecture. The feature representation learnt by the
convolution layers of ResNet with different batche sizes also
contributed to greater accuracy. Another significant reason
is the inclusion of local features, by locating the maximum
connected component on the binary map. The local area is
cropped from the input image (preprocessed image), and it
local features. In view of this, the global features extracted
by AlexNet and the local features generated by ResNet were
concatenated and fed into the fully-connected layer for final
ensembled classification model that forms MSDNet. The
results of this were evident, as the proposed MSDNet mod-
els outperformed both ResNet and AlexNet, emphasizing
the effectiveness of the global+local feature representations
towards anomaly classification.
The models performed best for the Elbow, Forearm,
Humerus, Wrist and Chest classes achieving >80% accuracy
which can be seen in Table 4, while, they showed satisfactory
results for the Finger, Hand and Shoulder classes. The other
two metrics, sensitivity and specificity, capture some addi-
tional aspects of the classification performance. The observed
sensitivity scores for the Elbow, Forearm, Hand, Wrist and
Chest classes were in the range of 83% to 89%, indicating
that the abnormal samples were correctly classified for these
classes to a larger extent. However, the specificity scores of
classes like Forearm and Hand indicate that the percentage of
normal scans correctly classified as normal was lower than that
of the other classes. The lowest sensitivity score was observed
for the shoulder class, indicating that the model could not very
well distinguish between normal and abnormal images, thus
requiring more detailed scrutiny and analysis. On average,
the proposed approach achieved an accuracy rate of 82.69%,
with sensitivity and specificity scores of 81.79% and 76.62%,
respectively, which indicates good classification performance.
As discussed earlier, the average Kappa statistic value
was 0.746, which indicates a substantial agreement on the
test samples with the expected values. However, radio-
graph findings are often judged subjectively, hence we also
used other metrics like accuracy, specificity and sensitiv-
ity, to gain more refined insights into the proposed model’s
performance. A graphical plot that illustrates the diag-
nostic ability of a classifier system in terms of the Area
under the ROC1 curve (AUC) is illustrated in Fig. 5. An
1
  Receiver Operating Characteristic.

13
MSDNet: a deep neural ensemble model for abnormality detection and classification of plain…

Table 5  Proposed model’s Classes Rajpurkar et al. Saif et al. (2019)∗ Banga and Solovyova and MSDNet
performance w.r.t Kappa Score (2017a) Waiganjo (2019) Solovyov (2020) (pro-
against State-of-the-art models posed)

Elbow 0.710 0.733∗ 0.617 0.715 0.736

Finger 0.389 0.735∗ 0.653 0.395 0.671
Forearm 0.737 0.785∗ 0.695 0.730 0.792
Hand 0.851 0.835∗ 0.584 0.862 0.754
Humerus 0.600 0.754∗ 0.599 0.602 0.676
Shoulder 0.729 0.855∗ 0.659 0.735 0.731
Wrist 0.931 0.907∗ 0.740 0.942 0.855
Chest – – – – 0.758

Approach marked with * used 50% of the data

Table 6  Benchmarking proposed model against state-of-the-art models using standard metrics

Models Layers Accuracy Sensitivity Specificity Kappa

MSDNet (Proposed) 25 0.82 0.81 0.76 0.74

DenseNet-169 (Chada 2019) ∗
169 0.79 0.72 0.88 0.60
DenseNet-201 (Chada 2019)∗ 201 0.82 0.81 0.84 0.64
InceptionResNetV2 (Chada 2019)∗ 164 0.82 0.81 0.83 0.64
EnsembleD [Dense, MobileN] (Banga and Waiganjo 2019) > 0.83 0.92 0.73 0.66
EnsembleE [Dense, Xcep] (Banga and Waiganjo 2019) > 0.71 0.77 0.63 0.41
MobileNet (Single) (Banga and Waiganjo 2019) 88 0.67 0.73 0.61 0.34
EnsembleD [MobileN, Xcep] (Banga and Waiganjo 2019) > 0.65 0.73 0.56 0.29

Note: Approach marked with * used Finger and Humerus class only. Average of two class results is pre

The classified images are fed into the proposed abnormal

Fig. 5  AUROC performance of the proposed ensemble model
AUC value of 0.9038 was obtained, indicating good per-
formance in distinguishing anomalous and non-anomalous
scans.
area detection process (as illustrated in Algorithm 1). The
first phase of the algorithm is to plot the histogram of the
given image i (Fig. 6b), the last peak value along the x-axis
of the histogram value is used to find the two largest objects/
blobs on the image (Fig. 6c). The connected components
of these two blobs are obtained (Fig. 6d), after which the
centroid points are captured from the convex hull masked
image (Fig. 6e). Using the centroid points a bounding box is
marked around the abnormal region of the image (Fig. 6f).
Abnormalities like fractures and cracks, or even hardware
artifacts used for setting bones like metal inserts and screws
are automatically detected and boundary region was marked
using the abnormal detection algorithm. During experimen-
tal validation, it was noted that among a total number of 538
abnormal case images, the algorithm correctly identified 445
images.
During the diagnostic report retrieval phase, for each
observation in Y (Test image features) to all the observa-
tions in X (Training image features), the pairwise distance
is measured using Cosine similarity and Standard Euclidean
distance measures. It was observed the overall retrieval accu-
racy was 78.03% and 77.26% with reference to the two tech-
niques. Due to its comparatively superior performance, we

13

Fig. 6  Illustration of the abnormal area detection process for sample images from the Shoulder class
Fig. 7  Sample of the model
generated report, along with the
ground-truth data
used Cosine similarity as the distance measure. When a test
image is fed into the model, the Cosine similarity is com-
puted to find the nearest match with the Indication, Findings
& Impression parameters. A sample report that is retrieved
by the model for a given test image is shown in Fig. 7.
The proposed model was benchmarked against state-
of-the-art models using Kappa score, the results of which
are tabulated in Table 5. It can be observed from the table
that the proposed MSDNet model showed substantial
improvement over other models. However, it underper-
formed slightly for the Hand and Wrist classes when com-
pared to Rajpurkar et al. (2017a)’s model. They used only
13
K. Karthik, S. Sowmya Kamath
one evaluation metric, i.e., Kappa, hence, it is difficult
to analyze the result because other standard evaluation
metrics like Accuracy, Sensitivity and Specificity were
not reported for their model. However, the deeper layers
might have contributed more to those specific classes as is
evidenced in the better kappa scores. Several other state-
of-the-art models were considered for the benchmarking
experiments, and were compared with the proposed model,
the results of which are presented in Table 6. Our model
outperformed all these state-of-the-art models (in terms of
Kappa score), proving its dominance in the abnormality
prediction clinical task.
MSDNet: a deep neural ensemble model for abnormality detection and classification of plain…
5 Conclusion and future work
In this paper, an ensemble deep neural model, MSDNet, that
combines global and local features for the clinical task of
radiological abnormality detection and classification. The
model is built on an abnormal region detection algorithm
which is used for identifying the anomalous regions in the
image. As the scan images vary in resolution and the type
of studies available in each class is different, the proposed
deep CNN model was trained separately for each of the
classes to observe its performance. Experimental validation
was observed through metrics like accuracy, sensitivity,
specificity and Kappa statistic score. The proposed model
achieved an accuracy of 82.69%, along with promising sen-
sitivity, specificity and Kappa statistic value of 0.746, indi-
cating good performance. The overall computation time was
also reduced during training, as our proposed model uses a
comparatively shallow architecture with 25layers compared
to the other state-of-the-art models. These models used a
much deeper architecture; however, the proposed model
outperformed them being less computationally expensive
due to shorter training time. We also designed an automated
diagnostic text report generation algorithm, which further
extended the proposed model pipeline for the identified
abnormal images, to alleviate radiologists’ cognitive burden
and improve the overall efficiency of the diagnosis.
Our extensive experiments revealed that, the proposed
approach often under-performed in the case of some
unbalanced classes, specifically, classes like Hand and
Wrist, despite their relatively distinctive features. We
intend to explore this issue in depth and further analyze it
for enhancing the designed model. In the datasets, a few
images with hardware artifacts were wrongly predicted
as non-anomalous, requiring further investigation. Differ-
ent image enhancement techniques will be experimented
with, for addressing typical imaging process faults like
blurring, over/under-exposure. Adopting super-resolution
techniques to improve scan quality of radiographic images,
for boosting abnormal region identification performance
is also another avenue we intend to pursue.
Declarations  pages 630–645. Springer
Conflict of Interest  The authors declare that they have no conflict of
interest. This article does not contain any studies with human partici-
pants or animals performed by any of the authors and confirms that no
ethical approval is required.
Data Availability Statement (DAS)  The datasets that were used and ana-
lyzed as part of the current study are openly and publicly available as:
MURA datasethttps://​stanf​ordml​group.​github.​io/​compe​titio​ns/​mura/
and Indiana Datasethttps://​www.​kaggle.​com/​raddar/​chest-​xrays-​india​
na-​unive​rsity.
References
Aowal MA, Minhaz AT, Ashraf K (2017) Abnormality detection
and localization in chest x-rays using deep convolutional neural
networks. arXiv preprint arXiv:​1705.​09850
Baadjou VAE, Roussel NA, Verbunt JAMCF, Smeets RJEM, de Bie
RA (2016) Systematic review: risk factors for musculoskeletal
disorders in musicians. Occup Med 66(8):614–622
Banga D, Waiganjo P (2019) Abnormality detection in musculoskel-
etal radiographs with convolutional neural networks (ensem-
bles) and performance optimization. arXiv preprint arXiv:​1908.​
02170
BMUS (2014) United States Bone and Joint Initiative: The Burden of
Musculoskeletal Diseases in the United States (BMUS), Fourth
Edition. http://​w ww.​b onea​n djoi​n tbur​d en.​o rg/​2 014-​r eport.
[Accessed on 1 May 2019]
Chada G (2019) Machine learning models for abnormality detec-
tion in musculoskeletal radiographs. Rep Med Cases Images
Videos 2(4):26
Cheng C-T, Tsung-Ying H, Tao-Yi L, Chih-Chen C, Ching-Cheng C,
I-Fang, Chien-Hung Liao, (2019) Application of a deep learn-
ing algorithm for detection and visualization of hip fractures on
plain pelvic radiographs. Eur Radiol 29(10):5469–5477
Chung SW, Han SS, Lee JW, Kyung-Soo O, Kim NR, Yoon, et al
(2018) Automated detection and classification of the proximal
humerus fracture by using deep learning algorithm. Acta Orthop
89:4
Dalia Y, Bharath A, Mayya V, Sowmya Kamath S (2021) Deepoa:
Clinical decision support system for early detection and severity
grading of knee osteoarthritis. In 2021 5th International Con-
ference on Computer, Communication and Signal Processing
(ICCCSP), pages 250–255. IEEE
Demner-Fushman D, Kohli MD, Rosenman MB, Shooshan SE, Rod-
riguez L, Antani S, Thoma GR, McDonald CJ (2016) Prepar-
ing a collection of radiology examinations for distribution and
retrieval. J Am Med Inform Assoc 23(2):304–310
Faes L, Wagner SK, Dun Jack F, Liu X et al (2019) Automated deep
learning design for medical image classification by health-care
professionals with no coding experience: a feasibility study.
Lancet Digit Health 1(5):e232–e242
Gale W, Oakden-Rayner L, Carneiro G, Bradley AP, Palmer LJ (2017)
Detecting hip fractures with radiologist-level performance using
deep neural networks. arXiv preprint arXiv:​1711.​06504
García-Floriano A, Ferreira-Santiago Á, Camacho-Nieto O, Márquez
C (2019) A machine learning approach to medical image clas-
sification: detecting age-related macular degeneration in fundus
images. Comput Electr Eng 75:218–229
Harzig P, Chen YY, Chen F, Lienhart R (2019) Addressing data bias
problems for chest x-ray image report generation. arXiv preprint
arXiv:​1908.​02123,
He K, Zhang X, Ren S, Sun J (2016) Identity mappings in deep
residual networks. In European conference on computer vision,
Ioppolo F, Rompe JD, Furia JP, Cacchio A (2014) Clinical applica-
tion of shock wave therapy (swt) in musculoskeletal disorders.
Eur J Phys Rehabil Med 50(2):217–30
Karthik K, Sowmya Kamath S (2021) A deep neural network model
for content-based medical image retrieval with multi-view clas-
sification. Vis Comput 37(7):1837–1850
Karthik K, Sowmya Kamath S (2021) Automated view orientation
classification for x-ray images using deep neural networks.
Smart computational intelligence in biomedical and health
informatics. CRC Press, Boca Raton, pp 61–72
Katara K, Sowmya K et al (2021) Deep neural models for automated
multi-task diagnostic scan management-quality enhancement,
13

view classification and report generation. Biomed Phys Eng
Express 8:1
Kim DH, MacKinnon T (2018) Artificial intelligence in fracture detec-
tion: transfer learning from deep convolutional neural networks.
Clin Radiol 73(5):439–445
Kitamura G, Chung CY, Moore BE (2019) Ankle fracture detection
utilizing a convolutional neural network ensemble implemented
with a small sample, de novo training, and multiview incorpora-
tion. J Digit Imaging 32(4):672–677
Krizhevsky A, Sutskever I, Hinton GE (2012) Imagenet classification
with deep convolutional neural networks. Adv Neural Inf Process
Syst 2:1097–1105
Krogue Justin D, Cheng Kaiyang V, Hwang Kevin M, Paul T et al
(2020) Automatic hip fracture identification and functional sub-
classification with deep learning. Radiol Artif Intell 2:2
Krupinski EA, Berbaum KS, Caldwell RT, Schartz KM, Kim J (2010)
Long radiology workdays reduce detection and accommodation
accuracy. J Am Coll Radiol 7:9
Kumar A, Kim J, Lyndon D, Fulham M, Feng D (2016) An ensemble
of fine-tuned convolutional neural networks for medical image
classification. IEEE J Biomed Health Inform 21:1
Mandikal V, Anantharaman A, Suhas BS (2019) An approach for mul-
timodal medical image retrieval using latent dirichlet allocation.
In Proceedings of the ACM India Joint International Conference
on Data Science and Management of Data, pages 44–51
Mayya V, Karthik K, Sowmya KS, Karadka K, Jeganathan J (2021)
Coviddx: Ai-based clinical decision support system for learning
covid-19 disease representations from multimodal patient data. In
HEALTHINF, pages 659–666
McHugh ML (2012) Interrater reliability: the kappa statistic. Biochem
Med Biochem Med 22(3):276–282
Mukesh BR, Harish T, Mayya V, Sowmya Kamath S (2021) Deep
learning based detection of diabetic retinopathy from inexpensive
fundus imaging techniques. In 2021 IEEE International Confer-
ence on Electronics, Computing and Communication Technolo-
gies (CONECCT), pages 1–6. IEEE
Nedumkunnel IM, George LE et al (2021) Explainable deep neural
models for covid-19 prediction from chest x-rays with region of
interest visualization. In 2021 2nd International Conference on
Secure Cyber Computing and Communications (ICSCCC), pages
96–101. IEEE
13
K. Karthik, S. Sowmya Kamath
Rajpurkar P, Irvin J, Bagul A, Ding D, Duan T et al (2017) Mura data-
set: Towards radiologist-level abnormality detection in musculo-
skeletal radiographs. arXiv preprint arXiv:​1712.​06957
Rajpurkar P, Irvin J, Zhu K, Yang B, Mehta H et al (2017) Chexnet:
Radiologist-level pneumonia detection on chest x-rays with deep
learning. arXiv preprint arXiv:​1711.​05225
Saif AFM, Shahnaz C, Zhu WP, Omair AM (2019) Abnormality detec-
tion in musculoskeletal radiographs using capsule network. IEEE
Access 7:81494–81503
Silvian SP, Maiya A, Resmi AT, Page T (2011) Antecedents of work
related musculoskeletal disorders in software professionals. Int J
Enterprise Netw Manag 4(3):247–260
Solovyova A, Solovyov I (2020) X-ray bone abnormalities detection
using mura dataset. arXiv preprint arXiv:​2008.​03356
Soundalgekar P, Kulkarni M, Nagaraju D (2018) Medical image
retrieval using manifold ranking with relevance feedback. In
2018 IEEE 12th International Conference on Semantic Comput-
ing (ICSC). IEEE,
Tataru C, Yi D, Shenoyas A, Ma A (2017) Deep learning for abnormal-
ity detection in chest x-ray images
Wærsted M, Hanvold TN, Veiersted KB (2010) Computer work and
musculoskeletal disorders of the neck and upper extremity: a sys-
tematic review. BMC Musculoskelet Disord 11(1):79
Woolf AD, Pfleger B (2003) Burden of major musculoskeletal condi-
tions. Bull World Health Organ 81:2
Yahalomi F, Chernofsky M, Werman M (2019) Detection of distal
radius fractures trained by a small set of x-ray images and faster
r-cnn. In Intelligent Computing-Proceedings of the Computing
Conference, pages 971–981. Springer,
Ying J, Dutta J, Guo N, Chenhui H, Zhou D, Sitek A, Li Q (2016)
Classification of exacerbation frequency in the copdgene cohort
using deep learning with deep belief networks. IEEE J Biomed
Health Inform 2:2
Publisher's Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.