1 of 19

Deep Transfer Learning Based Musculoskeletal

Abnormality Detection

Paper ID: 83

Authors

Abu Zahid Bin Aziz
Md. Al Mehedi Hasan

Dept. of Computer Science & Engineering

Rajshahi University of Engineering & Technology

3. Jungpil Shin

School of Computer Science and Engineering

University of Aizu, Japan.

Presented By

Abu Zahid Bin Aziz

Dept. of Computer Science & Engineering

Rajshahi University of Engineering & Technology

IJCACI 2020

International Joint Conference on Advances in Computational Intelligence

2 of 19

Outline

Introduction
Motivation
Literature Review
Dataset description
Data preprocessing
Architecture of our CNN model
Method evaluation metrics
Hyperparameter tuning
Results
Future scope
References

3 of 19

Introduction

The musculoskeletal system (locomotor system) is a human body system that provides our body with movement, stability, shape, and support.
When the musculoskeletal system is disrupted or affected through injury, it is called Musculoskeletal Disorders or MSDs.
Musculoskeletal conditions affect more than 1.7 billion people around the world [1].

https://o.quizlet.com/C54OYOCl.Z8XsNxtOUlywA_b.png

4 of 19

Introduction(Contd.)

Its symptoms include pain, joint noises, decreased range of motion etc.
In this paper, we proposed a transfer learning based approach for for detecting musculoskeletal abnormalities.

https://www.chiro-med.ca/wp-content/uploads/2016/04/23893421_s.jpg

5 of 19

Motivation

Doctors usually use laboratory and radiographic tests to diagnose MSDs.
Laboratory tests generally include blood tests to check:

ESR (erythrocyte sedimentation rate), level of creatine kinase, cyclic citrullinated peptide antibody, HLA-B27 etc.

Radiographic tests include:

x-ray, computed tomography (CT), magnetic resonance imaging (MRI), bone scanning etc.

But there are many obstacles in the way of effective diagnosis such as :

Surgical burdens in low- and middle-income countries, diagnostic errors by misreading radiographs etc.

These problems encouraged the increasing use of computational methodologies in recent years.

6 of 19

Literature Review

Rajpurkar et al. provided a comprehensive collection of musculoskeletal radiographs in their work. (MURA dataset) [2]

They also provided transfer learning based methodology as baseline.

Pradhan et al. used this dataset for human bone classification in their work [3].

They did not take account of abnormalities in their work.

Shubhajit et al. employed a deep CNN for detecting abnormalities [4].

Although their work produced decent accuracies in the testing datasets, they didn’t categorize their findings according to the study types

7 of 19

Dataset Description

We considered five study types (Elbow, Finger, Forearm, Humerus and Shoulder) from the MURA dataset
The ratio of training and testing datasets was about 10:1 in terms of the number of images.

Table 1. A summary of our dataset with respect to the number of images

Study Type	Training Dataset		Testing Dataset
Study Type	Abnormal Images	Normal Images	Abnormal Images	Normal Images
Elbow	2006	2625	230	235
Finger	1968	3138	247	214
Forearm	661	1164	151	150
Humerus	599	673	140	148
Shoulder	4168	4211	278	285
Total	9402	11811	1046	1032

8 of 19

Data Preprocessing

Our data preprocessing had three steps:

Data augmentation:

Horizontal flip
Random contrast
Random gamma
Random brightness
Shift scale rotate

Resizing & cropping:

First resize to (300,300,3)
The center crop to (224,224,3)

Normalization

9 of 19

Data Preprocessing(Contd.)

Fig 1. Data preprocessing steps for a positive study of elbow

10 of 19

Architecture of Our CNN Model

Fig 2. Architecture of our model.

11 of 19

Method Evaluation Metrics

We considered the following evaluation metrics for our model:

Sensitivity (SN)
Specificity (SP)
Accuracy (ACC)
Matthew’s correlation coefficient (MCC)

We also calculated another metric to compare with the existing work which is:

Cohen’s kappa statistic

12 of 19

Hyperparameter Tuning

In this work, we tuned three hyperparameters:

Pre-trained model:

ResNet50
InceptionV3
DenseNet169

We chose DenseNet169 because of its increased number of direct connections among convolution layers

Optimizers:

Adam,
RMSprop
SGD (Stochastic Gradient Descent)
Adagrad

We chose RMSprop because it produced the highest accuracy

13 of 19

Hyperparameter Tuning(Contd.)

Learning Rate:

We applied a learning rate scheduler on plateau which started from 0.0001

Fig 3. Changes of values of the learning rate with respect to the number of epochs.

14 of 19

Results

Study Type	ACC(%)	SN(%)	SP(%)	MCC
Elbow	87.18	82.74	93.77	0.750
Finger	82.5	75.68	90.31	0.621
Forearm	87.06	80.83	93.33	0.688
Humerus	85.76	87.14	84.46	0.715
Shoulder	86.56	86.49	86.64	0.695

Table 2. Performance of our model on different evaluation metrics

15 of 19

Results(Contd.)

Study Type	Rajpurkar et al.’s model(Cohen’s kappa)[2]	Our Model(Cohen’s kappa)	Improvement
Elbow	0.710	0.756	+6.47 %
Finger	0.389	0.650	+67.05 %
Forearm	0.737	0.732	-0.67 %
Humerus	0.600	0.715	+19.16 %
Shoulder	0.729	0.731	+0.27 %

Table 3. Comparison between our model and existing model using Cohen’s kappa statistic

16 of 19

Results(Contd.)

Fig 4. The performance of our model using the ROC curve.

17 of 19

Future Scopes

Providing a web application to facilitate users for academic purposes.
Looking for other study types to apply our methodology
Investigating different ways to utilize pre-trained models.
Investigating model’s performance by tuning other hyperparameters.
Providing a single classifier for all study types

18 of 19

References

[1] Woolf, A.D., Pfleger, B.: Burden of major musculoskeletal conditions. Bulletin of the world health organization 81, 646–656 (2003)

[2] Rajpurkar, P., Irvin, J., Bagul, A., Ding, D., Duan, T., Mehta, H., Yang, B., Zhu, K., Laird, D., Ball, R.L., et al.: Mura: Large dataset for abnormality detection in musculoskeletal radiographs. arXiv preprint arXiv:1712.06957 (2017)

[3] Pradhan, N., Dhaka, V.S., Chaudhary, H.: Classification of human bones using deep convolutional neural network. In: IOP Conference Series: Materials Science and Engineering. vol. 594, p. 012024. IOP Publishing (2019)

[4] Panda, S., Jangid, M.: Improving the model performance of deep convolutional neural network in mura dataset. In: Smart Systems and IoT: Innovations in Computing, pp. 531–541. Springer (2020)