GAN-based Medical Image Augmentation for Improving CNN Performance in Myositis Ultrasound Image Classification

Idiopathic inflammatory myopathies (IIMs) are a clinically heterogeneous group of immune-mediated inflammatory muscle diseases characterized by muscle weakness and multisystem involvement. Early diagnosis of IIMs can contribute to better prognosis and clinical management. The classifier based on a deep learning framework has proven to be an effective and noninvasive technique for classifying IIMs from ultrasound images. However, as the availability of clinical ultrasound data for IIMs is always in paucity, data augmentation is indispensable to improve the classification accuracy. To this end, this work proposes the use of generative adversarial networks (GAN) with attentional mechanism to generate IIMs ultrasound data. We employ VGG16 with transfer learning as a classification network and find that the proposed method improves by 1.10% compared to no data enhancement, reaching 92.23% on the validation dataset. It is also shown that with the same classification network, the data augmentation by GAN is more competitive in facilitating classification compared to the traditional data augmentation.