Transfer Learning with Fine-Tuning on MobileNet and GRAD-CAM for Bones Abnormalities Diagnosis

Abstract

Osteoarthritis is a common medical condition. Unfortunately, despite the support of X-ray imaging technology in diagnosis, the accuracy of diagnostic results still depends on human factors. Furthermore, when errors do occur, they are often detected late, leading to a waste of time, money, and even disability for the patient. This study has deployed and evaluated transfer learning techniques in abnormal and normal bone images classification on X-ray images collected from the dataset of MUsculoskeletal RAdiographs (MURA) with 17,367 images and then leveraged techniques for results explanations of learning algorithms such as Gradient-weighted Class Activation Mapping (GRAD-CAM) to provide visual highlighted interesting areas in the images which can be signals for anomalies in bones. The classification performance using MobileNet with techniques of hyper-parameters fine-tuning can reach an accuracy of 0.84 in abnormal and normal bone classification tasks on the wrist, humerus, and elbow. The work is expected to provide visual support for doctors in diagnosing and identifying bone anomalies on X-ray images based on leveraging advancements from Artificial Intelligence techniques for medical imaging analysis.