Abnormality Detection in Musculoskeletal Radiographs using EfficientNets

Abstract

Abnormality detection in musculoskeletal radiographs, a regular task for radiologists, requires both experiences and efforts. To increase the number of radiographs interpreted each day, this paper presents cost-efficient deep learning models based on ensembles of EfficientNet architectures to help automate the detection process. We investigate the transfer learning performance of ImageNet pre-trained checkpoints on the musculoskeletal radiograph (MURA) dataset which is very different from the ImageNet dataset. The experimental results show that, the ImageNet pre-trained checkpoints have to be retrained on the entire MURA training set, before being trained on a specific study type. The performance of the EfficientNet-based models is shown to be superior to three baseline models. In particular, EfficientNet-B3 not only achieved the overall Cohen's Kappa score of 0.717, compared to the scores 0.680, 0.688, and 0.712 for MobileNetV2, DenseNet-169, and Xception, respectively, but also being better in term of efficiency.