Musculoskeletal abnormality detection on x-ray using transfer learning

Abstract

Radiographic studies are a common technique employed to detect a variety of diseases, in which the detection of musculoskeletal abnormalities has proven to be a crucial task. This thesis proposes the use of deep learning techniques to detect musculoskeletal abnormalities in the MURA dataset, one of the largest collections of radiographic studies. In particular, we use transfer learning techniques such as feature extraction and fine-tuning to well-known models for visual tasks such as InceptionV3, VGG and SqueezeNet, among others. Additionally, we present a tool based on class activation mappings to aid in visualizing the decision of our models. The results obtained show that transfer learning techniques can be applied to deep convolutional neural networks pre-trained on non-medical images, while achieving comparable results to the state-of-the-art.