Data Augmentation Transformations for Self-Supervised Learning with Ultrasound

Abstract

Central to joint embedding self-supervised learning is the choice of data augmentation pipeline used to produce positive pairs. This study developed and investigated data augmentation strategies for medical ultrasound. Three pipelines were studied: BYOL augmentations (as a baseline), AugUS-v1 – a pipeline designed to retain semantic content, and AugUS-v2 – a pipeline designed from baseline and AugUS-v1 transformations. Evaluation of SimCLR-pretrained models on diagnostic downstream tasks in lung ultrasound yielded mixed results. The use of AugUS-v1 led to the best performance on COVID-19 classification on a public dataset. However, BYOL and AugUS-v2 outperformed AugUS-v1 on A-line versus B-line classification. AugUS-v2 decidedly obtained the greatest performance on pleural effusion detection. The salient findings were that ultrasound-specific transformations may be suitable for some tasks more than others, and that the random crop and resize transformation was instrumental for all tasks.