Professional Documents
Culture Documents
KK SKS J5 MSDNetADeepNeuralEnsembleModel
KK SKS J5 MSDNetADeepNeuralEnsembleModel
https://doi.org/10.1007/s12652-022-03835-8
ORIGINAL RESEARCH
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
13
Vol.:(0123456789)
K. Karthik, S. Sowmya Kamath
thus requires handling a large amount of data, and the possi- pattern in the medical image. These patterns can be from
bility exists that the intensive labor may lead to misdiagnosis a collection of geometric shapes, like corners, edges, lines,
(Ying et al. 2016). A study carried out to measure radiolo- circles, contours, etc. These shapes are detected by means of
gist exhaustion during the process of interpreting musculo- a specific method or technique, for example, fracture lines
skeletal radiographs (Krupinski et al. 2010), reported that can be detected by various edge detection techniques. How-
a significant decrease in their overall efficiency during the ever, these can be obscured in some scan view orientations,
process of fracture detection from diagnostic scans, at the affecting the detection performance Karthik and Kamath
end of the work day when compared to the beginning. (2021b). Further, the performance of detecting the abnor-
Musculo-skeletal Disorders (MSDs) are often debilitat- malities can be improved by integrating prior knowledge
ing and affect a person’s daily activities to a large extent. of abnormalities, such as the location, size or orientation.
Hairline fractures, carpal tunnel syndrome, ligament tear are With the advancement of applications of artificial intelli-
some examples of such disorders, the potential causes of gence (AI) based models for healthcare, research in Machine
which could be work-related accidents, workplace fatigue, learning (ML) and deep learning (DL) have shown great
sport injuries etc. The daily routines of busy professionals promise in clinical tasks like disease diagnosis (Mukesh
like those who do repetitive tasks on computers (banking, et al. 2021; Katara et al. 2021), medical image classification
software professionals etc), have seen increasing occur- (Faes et al. 2019; Kumar et al. 2016; García-Floriano et al.
rence of neck and lower back pain. Researchers that studied 2019) and image retrieval (Mandikal et al. 2019; Soundal-
occupation-related cases of such professionals observed that gekar et al. 2018; Karthik and Kamath 2021a) among oth-
physical and individual factors affected the health of the ers. The performance of these systems in clinical settings
worker’s musculo-skeletal system over time (Silvian et al. can revolutionise the way healthcare services are delivered,
2011; Wærsted et al. 2010). Playing Related Musculo-skel- especially in a labor-intensive field like radiography, where,
etal Disorder (PRMD) is a common ailment affecting career the radiologist is expected to manually check each scan and
musicians, it has been reported that string quarter players write a list of observations, for enabling diagnosis by the
experience more MSDs than other instrumentalists (Baadjou referring doctors.
et al. 2016). Early diagnosis can help and emerging diagnos- Image processing and Computer Vision (CV) based
tic procedures like physiotherapy, shock wave therapy etc techniques have been applied for designing applications
have been used to treat any mild to medium symptoms of for surgical and imaging interventions. Such systems are
pain caused by MSDs. Primarily, such treatments can help extend clinical decision making capabilities to the healthcare
give some relief in disorders like tendonitis and osteoarthri- professionals, by automating certain tasks related to diagno-
tis (which cause pain, swelling in the elbow, wrist, finger, sis, or by forecasting the severity of several abnormalities
thigh, and other parts of the body, due to overuse) and bone (Nedumkunnel et al. 2021; Mayya et al. 2021). Incorporat-
fractures (Ioppolo et al. 2014). ing AI in these systems to support learning behaviour so
Early detection and effective diagnosis for assessing any that systems can detect abnormalities at the earliest disease
abnormalities in scans, is thus crucial for further treatment. onset in a wide variety of diagnostic media are of critical
Currently, such abnormality detection is predominantly a importance. The radiologist can utilize these insights for
manual task performed by expert radiologists or orthopae- enabling and optimizing the quality of diagnosis (Katara
dists, making it a time-consuming process demanding sig- et al. 2021). The marked region can help the physicians to
nificant time and effort from trained healthcare profession- focus on early and effective treatment recommendations.
als. Given the large volume of diagnostic scans undergone Further, automated retrieval of radiological diagnosis reports
by patients each day in hospitals, this is a significant burden will minimize the manual work involved in reporting obser-
on available healthcare resources, often delaying treatment vations from the radiological images, while also alleviat-
(Dalia et al. 2021). In view of this, automated algorithms for ing the cognitive burden of the radiologists due to the huge
classification and detection of anomalies is a crucial require- load of cases that they typically handle each day. The major
ment and can revolutionize the ways in which healthcare can contributions of the research work presented in this article
be delivered at the point-of-care. are listed below:
Abnormalities that exist in medical images can be char-
acterized and detected in various ways. A traditional and 1. Design of ensemble deep neural model for classifying
common categorization is using a shape model. Certain abnormal radiography images.
abnormalities have unique geometric shapes, and their sizes 2. Design of a boundary detection algorithm for identifying
are helpful for detection of disease staging and prognosis. the region of interest for facilitating anomaly detection.
Considering an example, some tumor regions are round in 3. A content-based medical report retrieval approach for
shape and tend to grow in size. In such cases, detection of automatically generating diagnostic text reports for a
abnormalities can be considered as a task with particular
13
MSDNet: a deep neural ensemble model for abnormality detection and classification of plain…
given image based on the anomalies identified by the detection using musculoskeletal radiographs. Their model’s
proposed models. performance was not satisfactory as its Cohen Kappa score
was lesser than that of the DenseNet169 model (Rajpurkar
The rest of this paper is organized as follows: In Sect. 2, we et al. 2017a), but in terms of F1 score, it outperformed the
present a detailed discussion on available recent work in the DenseNet model. Solovyova and Solovyov (2020) proposed
area of interest. Section 3 details the proposed methodology a 169 layered Deep CNN model for abnormal X-ray image
for abnormality classification, localization and text retrieval classification. To increase the quality of the initial pre-pro-
for different radiograph classes. The details of experiments cessing for identifying the region of interest, they used dif-
conducted for benchmarking the proposed approaches are ferent augmentation techniques. Their model performed well
presented in Sect. 4, followed by conclusion and future work. on bone abnormality classification, however, it is limited to
the binary classification problem, i.e., whether an anomaly
exists or not.
2 Related work An ensemble model built on Inception V3, ResNet,
and Xception was developed by Kitamura et al. (2019).
Recent advancements in machine learning and deep learn- The model’s performance was measured considering the
ing architectures have emphasized the potential improve- multiple views of the ankle radiographs, to determine the
ments in classification tasks, with major impact in the field effect of fracture detection. However, the model was trained
of diagnostic image analysis and management. The avail- and tested on a small dataset consisting of 596 normal
ability of large diagnostic image datasets have also helped and abnormal ankle cases. Chung et al. (2018) proposed
in the development of deep learning models for different a deep CNN after augmenting the training dataset to clas-
tasks such as image classification, recognition, segmentation sify proximal humerus fractures. They claimed that their
and so on. Efforts are in progress for developing large, open model showed superior performance when compared with
and standard repositories for facilitating medical image man- the human groups of general physicians and orthopedists.
agement. The MURA dataset made available by Rajpurkar However, their study neglected to address the clinical task
et al. (2017a) consists of an extensive collection of musculo- of fracture region detection, which is of primary importance.
skeletal studies of upper extremities like hand, wrist, elbow, Gale et al. (2017) proposed a hip fracture detection approach
finger, forearm, humerus and shoulder radiography classes. based on deep neural networks on a small annotated data-
The Indiana university dataset (Rajpurkar et al. 2017b) pro- set which includes 3,354 images, with 348 fractures, and
vides chest X-rays with textual reports reporting observed reported good results.
abnormal conditions. Yahalomi et al. (2019) proposed a Faster R-CNN based
Automatically deciding if a radiography study is normal model for identification of fractures in anteroposterior hand
or abnormal is a task that can significantly reduce the bur- X-ray images. Their model was reported to be more accurate
den of radiologists who handle a large volume and variety in detecting the fracture regions than by the radiologist’s
of radiological images each day. Rajpurkar et al. (2017a) findings, however, the model used a small set of diagnosed
developed a deep 169-layer convolutional neural network X-ray images and was pre-trained on the Imagenet dataset.
for predicting the abnormality of musculoskeletal studies. Kim and MacKinnon (2018) used Inception v3 as a trans-
The model predicts a study as abnormal if the prediction rate fer learning approach for deep CNNs for classifying wrist
was found to be greater than 0.5. Here, the variable-sized radiographs into fractured/no-fracture classes. However,
images were scaled to 320×320 and the network weights they used a very small dataset and thus their approach is not
were set according to the pre-trained ImageNet model with a scalable. Krogue et al. (2020) proposed a DenseNet deep
learning rate of 0.0001 and minibatch size of 8. Their results learning model for identification and classification of hip
performed best for wrist and humerus classes using Cohen’s fractures. The model considered a single class image for
kappa statistic with a 95% confidence interval compared prediction, unlike a human interpreter, who may look at sev-
with the three best radiologists. eral views for abnormality classification. Cheng et al. (2019)
Saif et al. (2019) developed a model based on the Capsnet used a deep CNN model to detect and visualize hip fractures.
architecture called a capsule network for identification of Further, their model used Grad-CAM models to check the
abnormalities present in musculoskeletal radiographs. By model’s prediction by highlighting the fracture region.
varying the image size to different pixel sizes, the perfor- Tataru et al. (2017) used three different kinds of CNN
mance of the network was observed at each cases. Also the models like GoogLeNet, InceptionV3 and ResNet, for
proposed capsnet model achieved an improved kappa score classifying chest X-rays as either normal, abnormal, or
of 10% when compared to the 169 layered Densenet model, extremely high-risk. The original image sizes were down-
while using 50% less training data. Banga and Waiganjo sampled to 512*512 and 224×224 from 3000×3000 pixels
(2019) proposed an ensembled 200 model for abnormality before feeding them to the neural model. Among the three,
13
K. Karthik, S. Sowmya Kamath
GoogleNet performed the best for the chest X-ray image extremity classes like Hand, Forearm, Wrist, Finger, Shoul-
classification task. Adding to this, detecting the abnor- der, Humerus and Elbow. The second dataset, Indiana Uni-
mal region or including the text reports already available versity dataset consists of chest X-ray images, along with
for chest X-rays in the generation of the reports could be indications, findings, and impressions in a textual form
added in their methodology pipeline. Aowal et al. (2017) for each image. These two datasets were together used for
employed DCNN based models for abnormality detection addressing the three different clinical tasks, each involving
in chest X-rays. Their approach showed consistent detection data belonging to two data modalities, X-ray images and
results even when the number of available training examples clinical text reports, thus resulting in multimodal data-
were low, and when random & multiple train-test data splits sets. Due to this, the proposed ensemble models are more
were used. However, the DCNN did not perform well across robust, capable of dealing with varied types of abnormali-
all abnormalities. Their study considered only frontal chest ties in underlying radiographic images. The three clinical
X-ray images and text reports were also not incorporated. tasks undertaken form a significant part of a typical clinical
Harzig et al. (2019) proposed a dual-word LSTM sentence workflow that is managed on a daily basis by a radiologist
generation model separately for abnormal and normal chest in hospital scenarios
X-ray images. They reported that dual-word LSTMs helped
increase the number of distinct sentences, but, failed to iden- 1. Classifying a given diagnostic image as either normal or
tify abnormal regions in the image. abnormal.
Table 1 summarizes the major research works related 2. Abnormal region detection for localize and visualizing
to abnormality classification, localization of the abnormal identified abnormal areas.
region and text report summary for a given image under 3. Automated diagnostic text report generation.
test. From the extensive study of existing works under-
taken, we observed that the problem of classification and The proposed methodology is illustrated in Fig. 1. The
identification of abnormalities still evince significant scope images are first pre-processed, after which they are fed into
for further research. Several challenges exist that need to the proposed neural network model for abnormality classi-
be addressed, introduced due to the multimodal nature of fication. For preprocessing, the images present in the data-
diagnostic images, i.e., the complexity of the image and the sets are of different sizes. The height and the width of each
extensive variations across different imaging modalities. image are not uniform, hence these are resized to a standard
After an image is classified as abnormal, identifying the size of 256 × 256 for AlexNet and 224 × 224 for ResNet
fracture region or detecting the area of abnormality has not to train our models efficiently. The input size was selected
received much attention. It was also observed that, diagnos- based on the input layer of the neural network to preserve
tic text report generation based on the observations obtained the information as much as possible and simultaneously to
from the underlying neural models is under-explored. In the reduce the overall computational cost during training the
overall process of radiography image examination, the inclu- model. Any abnormal regions are identified by the proposed
sion of the identified findings in the form of an automatically abnormal region detection algorithm by marking a bounding
generated textual report would be an added advantage for box around the identified region. Next, the task of automatic
physicians. retrieval of diagnosis report for the given image under test,
In view of these observations, our work encompasses is performed. The methodology adopted for addressing each
developing an ensemble deep neural network model for of these tasks are discussed in more detail in the subsequent
automated abnormality detection and classification. The sections. The proposed MSDNet ensemble model is the core
abnormalities present in the images are identified using our of the pipeline that is used for the three clinical tasks, which
developed abnormal region detection algorithm. Further, the is trained to extract features and image index values that will
features generated by the ensemble neural model are used for facilitate the classification, identification and localization of
the auto retrieval of radiological text reports, thus reducing anomalies, and also automated diagnostic text report genera-
the radiologist’s workload and making it more manageable. tion for a given input image.
3.1 Abnormality classification
3 Proposed approach
The architecture of the proposed MSDNet model is depicted
The proposed approach for abnormality classification, in Fig. 2. The proposed network is built as an ensemble of
abnormality localization and automated diagnostic report the AlexNet (Krizhevsky et al. 2012) and ResNet18 (He et al.
retrieval we used two publicly available standard datasets. 2016) architectures for initial classification of the category of
The first one, the MURA dataset (Rajpurkar et al. 2017a) the image, i.e. abnormal or normal. The global features are
provides musculoskeletal radiograph images of seven upper obtained from the first model by directly feeding the original
13
Table 1 Summary of key works on abnormality detection and report generation
Author & Year Methodology Merits Observed limitations
Rajpurkar et al. (2017a) Deep CNN model with 169 layers Abnormality prediction if threshold >0.5, binary Only Cohen’s kappa statistic used for evaluation
classification
Tataru et al. (2017) GoogLeNet,InceptionV3 and ResNet Visualization results demonstrate that macroscopic Abnormal region detection is not performed
features are learned effectively by the model
Aowal et al. (2017) 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
Kim and MacKinnon (2018) 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
Chung et al. (2018) 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
Saif et al. (2019) Capsule network architecture Performance of the network features vanquished No abnormality region detection
the limitations of CNN
Banga and Waiganjo (2019) Ensembled200 model The model outperformed DenseNet model Cohen Kappa’s score was not computed. No
bounding-box or abnormality region detection
Chada (2019) 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
Yahalomi et al. (2019) 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
Kitamura et al. (2019) 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
Cheng et al. (2019) DCNN and Grad-CAM Localizing fracture lesions Only four classes, Ankles, elbows, feet, and wrists,
were used in the study
Krogue et al. (2020) 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
Harzig et al. (2019) 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
Solovyova and Solovyov (2020) Pre-processing and deep 169 layer CNN based The model showed a good results in bone abnor- focused only on the binary classification problem
neural model malities detection
13
K. Karthik, S. Sowmya Kamath
13
MSDNet: a deep neural ensemble model for abnormality detection and classification of plain…
13
K. Karthik, S. Sowmya Kamath
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…
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
Preoperative renal transplant The lungs and pleural spaces show no acute abnormal- No acute pulmonary findings. Mild cardiomegaly
ity. Stable left upper lobe calcified granuloma. Heart
size is mildly enlarged, pulmonary vascularity within
normal limits. Mild tortuosity of the descending
thoracic aorta
Chest and midback pain 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
Shortness of breath 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
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)
Note: Approach marked with * used Finger and Humerus class only. Average of two class results is pre
13
K. Karthik, S. Sowmya Kamath
Fig. 6 Illustration of the abnormal area detection process for sample images from the Shoulder class
used Cosine similarity as the distance measure. When a test one evaluation metric, i.e., Kappa, hence, it is difficult
image is fed into the model, the Cosine similarity is com- to analyze the result because other standard evaluation
puted to find the nearest match with the Indication, Findings metrics like Accuracy, Sensitivity and Specificity were
& Impression parameters. A sample report that is retrieved not reported for their model. However, the deeper layers
by the model for a given test image is shown in Fig. 7. might have contributed more to those specific classes as is
The proposed model was benchmarked against state- evidenced in the better kappa scores. Several other state-
of-the-art models using Kappa score, the results of which of-the-art models were considered for the benchmarking
are tabulated in Table 5. It can be observed from the table experiments, and were compared with the proposed model,
that the proposed MSDNet model showed substantial the results of which are presented in Table 6. Our model
improvement over other models. However, it underper- outperformed all these state-of-the-art models (in terms of
formed slightly for the Hand and Wrist classes when com- Kappa score), proving its dominance in the abnormality
pared to Rajpurkar et al. (2017a)’s model. They used only prediction clinical task.
13
MSDNet: a deep neural ensemble model for abnormality detection and classification of plain…
In this paper, an ensemble deep neural model, MSDNet, that Aowal MA, Minhaz AT, Ashraf K (2017) Abnormality detection
and localization in chest x-rays using deep convolutional neural
combines global and local features for the clinical task of networks. arXiv preprint arXiv:1705.09850
radiological abnormality detection and classification. The Baadjou VAE, Roussel NA, Verbunt JAMCF, Smeets RJEM, de Bie
model is built on an abnormal region detection algorithm RA (2016) Systematic review: risk factors for musculoskeletal
which is used for identifying the anomalous regions in the disorders in musicians. Occup Med 66(8):614–622
Banga D, Waiganjo P (2019) Abnormality detection in musculoskel-
image. As the scan images vary in resolution and the type etal radiographs with convolutional neural networks (ensem-
of studies available in each class is different, the proposed bles) and performance optimization. arXiv preprint arXiv:1908.
deep CNN model was trained separately for each of the 02170
classes to observe its performance. Experimental validation BMUS (2014) United States Bone and Joint Initiative: The Burden of
Musculoskeletal Diseases in the United States (BMUS), Fourth
was observed through metrics like accuracy, sensitivity, Edition. http://w ww.b onean djoin tburd en.o rg/2 014-r eport.
specificity and Kappa statistic score. The proposed model [Accessed on 1 May 2019]
achieved an accuracy of 82.69%, along with promising sen- Chada G (2019) Machine learning models for abnormality detec-
sitivity, specificity and Kappa statistic value of 0.746, indi- tion in musculoskeletal radiographs. Rep Med Cases Images
Videos 2(4):26
cating good performance. The overall computation time was Cheng C-T, Tsung-Ying H, Tao-Yi L, Chih-Chen C, Ching-Cheng C,
also reduced during training, as our proposed model uses a I-Fang, Chien-Hung Liao, (2019) Application of a deep learn-
comparatively shallow architecture with 25layers compared ing algorithm for detection and visualization of hip fractures on
to the other state-of-the-art models. These models used a plain pelvic radiographs. Eur Radiol 29(10):5469–5477
Chung SW, Han SS, Lee JW, Kyung-Soo O, Kim NR, Yoon, et al
much deeper architecture; however, the proposed model (2018) Automated detection and classification of the proximal
outperformed them being less computationally expensive humerus fracture by using deep learning algorithm. Acta Orthop
due to shorter training time. We also designed an automated 89:4
diagnostic text report generation algorithm, which further Dalia Y, Bharath A, Mayya V, Sowmya Kamath S (2021) Deepoa:
Clinical decision support system for early detection and severity
extended the proposed model pipeline for the identified grading of knee osteoarthritis. In 2021 5th International Con-
abnormal images, to alleviate radiologists’ cognitive burden ference on Computer, Communication and Signal Processing
and improve the overall efficiency of the diagnosis. (ICCCSP), pages 250–255. IEEE
Our extensive experiments revealed that, the proposed Demner-Fushman D, Kohli MD, Rosenman MB, Shooshan SE, Rod-
riguez L, Antani S, Thoma GR, McDonald CJ (2016) Prepar-
approach often under-performed in the case of some ing a collection of radiology examinations for distribution and
unbalanced classes, specifically, classes like Hand and retrieval. J Am Med Inform Assoc 23(2):304–310
Wrist, despite their relatively distinctive features. We Faes L, Wagner SK, Dun Jack F, Liu X et al (2019) Automated deep
intend to explore this issue in depth and further analyze it learning design for medical image classification by health-care
professionals with no coding experience: a feasibility study.
for enhancing the designed model. In the datasets, a few Lancet Digit Health 1(5):e232–e242
images with hardware artifacts were wrongly predicted Gale W, Oakden-Rayner L, Carneiro G, Bradley AP, Palmer LJ (2017)
as non-anomalous, requiring further investigation. Differ- Detecting hip fractures with radiologist-level performance using
ent image enhancement techniques will be experimented deep neural networks. arXiv preprint arXiv:1711.06504
García-Floriano A, Ferreira-Santiago Á, Camacho-Nieto O, Márquez
with, for addressing typical imaging process faults like C (2019) A machine learning approach to medical image clas-
blurring, over/under-exposure. Adopting super-resolution sification: detecting age-related macular degeneration in fundus
techniques to improve scan quality of radiographic images, images. Comput Electr Eng 75:218–229
for boosting abnormal region identification performance Harzig P, Chen YY, Chen F, Lienhart R (2019) Addressing data bias
problems for chest x-ray image report generation. arXiv preprint
is also another avenue we intend to pursue. arXiv:1908.02123,
He K, Zhang X, Ren S, Sun J (2016) Identity mappings in deep
residual networks. In European conference on computer vision,
Declarations pages 630–645. Springer
Ioppolo F, Rompe JD, Furia JP, Cacchio A (2014) Clinical applica-
tion of shock wave therapy (swt) in musculoskeletal disorders.
Conflict of Interest The authors declare that they have no conflict of
Eur J Phys Rehabil Med 50(2):217–30
interest. This article does not contain any studies with human partici-
Karthik K, Sowmya Kamath S (2021) A deep neural network model
pants or animals performed by any of the authors and confirms that no
for content-based medical image retrieval with multi-view clas-
ethical approval is required.
sification. Vis Comput 37(7):1837–1850
Karthik K, Sowmya Kamath S (2021) Automated view orientation
Data Availability Statement (DAS) The datasets that were used and ana-
classification for x-ray images using deep neural networks.
lyzed as part of the current study are openly and publicly available as:
Smart computational intelligence in biomedical and health
MURA datasethttps://stanfordmlgroup.github.io/competitions/mura/
informatics. CRC Press, Boca Raton, pp 61–72
and Indiana Datasethttps://www.kaggle.com/raddar/chest-xrays-india
Katara K, Sowmya K et al (2021) Deep neural models for automated
na-university.
multi-task diagnostic scan management-quality enhancement,
13
K. Karthik, S. Sowmya Kamath
view classification and report generation. Biomed Phys Eng Rajpurkar P, Irvin J, Bagul A, Ding D, Duan T et al (2017) Mura data-
Express 8:1 set: Towards radiologist-level abnormality detection in musculo-
Kim DH, MacKinnon T (2018) Artificial intelligence in fracture detec- skeletal radiographs. arXiv preprint arXiv:1712.06957
tion: transfer learning from deep convolutional neural networks. Rajpurkar P, Irvin J, Zhu K, Yang B, Mehta H et al (2017) Chexnet:
Clin Radiol 73(5):439–445 Radiologist-level pneumonia detection on chest x-rays with deep
Kitamura G, Chung CY, Moore BE (2019) Ankle fracture detection learning. arXiv preprint arXiv:1711.05225
utilizing a convolutional neural network ensemble implemented Saif AFM, Shahnaz C, Zhu WP, Omair AM (2019) Abnormality detec-
with a small sample, de novo training, and multiview incorpora- tion in musculoskeletal radiographs using capsule network. IEEE
tion. J Digit Imaging 32(4):672–677 Access 7:81494–81503
Krizhevsky A, Sutskever I, Hinton GE (2012) Imagenet classification Silvian SP, Maiya A, Resmi AT, Page T (2011) Antecedents of work
with deep convolutional neural networks. Adv Neural Inf Process related musculoskeletal disorders in software professionals. Int J
Syst 2:1097–1105 Enterprise Netw Manag 4(3):247–260
Krogue Justin D, Cheng Kaiyang V, Hwang Kevin M, Paul T et al Solovyova A, Solovyov I (2020) X-ray bone abnormalities detection
(2020) Automatic hip fracture identification and functional sub- using mura dataset. arXiv preprint arXiv:2008.03356
classification with deep learning. Radiol Artif Intell 2:2 Soundalgekar P, Kulkarni M, Nagaraju D (2018) Medical image
Krupinski EA, Berbaum KS, Caldwell RT, Schartz KM, Kim J (2010) retrieval using manifold ranking with relevance feedback. In
Long radiology workdays reduce detection and accommodation 2018 IEEE 12th International Conference on Semantic Comput-
accuracy. J Am Coll Radiol 7:9 ing (ICSC). IEEE,
Kumar A, Kim J, Lyndon D, Fulham M, Feng D (2016) An ensemble Tataru C, Yi D, Shenoyas A, Ma A (2017) Deep learning for abnormal-
of fine-tuned convolutional neural networks for medical image ity detection in chest x-ray images
classification. IEEE J Biomed Health Inform 21:1 Wærsted M, Hanvold TN, Veiersted KB (2010) Computer work and
Mandikal V, Anantharaman A, Suhas BS (2019) An approach for mul- musculoskeletal disorders of the neck and upper extremity: a sys-
timodal medical image retrieval using latent dirichlet allocation. tematic review. BMC Musculoskelet Disord 11(1):79
In Proceedings of the ACM India Joint International Conference Woolf AD, Pfleger B (2003) Burden of major musculoskeletal condi-
on Data Science and Management of Data, pages 44–51 tions. Bull World Health Organ 81:2
Mayya V, Karthik K, Sowmya KS, Karadka K, Jeganathan J (2021) Yahalomi F, Chernofsky M, Werman M (2019) Detection of distal
Coviddx: Ai-based clinical decision support system for learning radius fractures trained by a small set of x-ray images and faster
covid-19 disease representations from multimodal patient data. In r-cnn. In Intelligent Computing-Proceedings of the Computing
HEALTHINF, pages 659–666 Conference, pages 971–981. Springer,
McHugh ML (2012) Interrater reliability: the kappa statistic. Biochem Ying J, Dutta J, Guo N, Chenhui H, Zhou D, Sitek A, Li Q (2016)
Med Biochem Med 22(3):276–282 Classification of exacerbation frequency in the copdgene cohort
Mukesh BR, Harish T, Mayya V, Sowmya Kamath S (2021) Deep using deep learning with deep belief networks. IEEE J Biomed
learning based detection of diabetic retinopathy from inexpensive Health Inform 2:2
fundus imaging techniques. In 2021 IEEE International Confer-
ence on Electronics, Computing and Communication Technolo- Publisher's Note Springer Nature remains neutral with regard to
gies (CONECCT), pages 1–6. IEEE jurisdictional claims in published maps and institutional affiliations.
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