Improving the Model Performance of Deep Convolutional Neural Network in MURA Dataset

Abstract

Deep convolutional neural networks have recently become one of the most powerful and expressive learning models for image pattern recognition and classification problems. And its wide use in the medical field has gathered immense success and importance in current scenarios. In this work, we made use of the Deep Convolutional Neural Network (CNN) over a large set of data called MURA (Musculoskeletal Radiographs Abnormality) and tried to improve the model performance (in terms of maximizing the accuracy and minimizing the loss) through the use of six different deep learning optimizers as well as the use of dropout regularizers (to avoid overfitting of data). The network architecture was trained using these optimizers to obtain the best possible model parameters that can easily exceed the previous level of performance in the classification tasks of abnormality detection in the human muscoskeletal system based on radiographs. Our model achieved the highest training accuracy of 95.98% and highest valid accuracy of 93.70%, which is better than the one achieved in the previous work. Similarly, in terms of train loss and valid loss also the model achieved the lowest loss of 0.11 and 0.22 respectively which is lower than the loss obtained in the previous work.