Self-Supervised Learning With BYOL for Non-Alcoholic Fatty Liver Disease Diagnosis Using Ultrasound Imaging

Abstract

Purpose:The study aims to evaluate the effectiveness of Bootstrap Your Own Latent (BYOL), a self-supervised learning method for diagnosing NAFLD from ultrasound images using limited labeled data, which represents a novel approach in this domain. Self-supervised learning provides an alternative approach to traditional supervised learning by learning useful representations from unlabeled data, thereby reducing the time and cost required by radiologists to annotate images.Methods:The pre-trained ResNet-50 and ResNet-101 on the labeled ImageNet dataset were used for BYOL pre-training on ultrasound images without relying on labels. The training was conducted using default and custom augmentation, as well as balanced and imbalanced class distribution protocols. The model was then evaluated using linear and fine-tuning protocols with varying percentages of labeled data. The model was trained using three shuffled subsets, each trained 10 times. The custom augmentation set was derived by testing various augmentation settings using 100% and 1% of the labels to enhance feature learning.Results:BYOL with ResNet-101 and using the proposed custom augmentation set achieved average accuracies of 93.44%, 92.29%, and 88.49% using 100%, 10%, and 1% of the training labels across three shuffled datasets. In addition, our proposed method attained an average accuracy of 96.9% using patient-specific leave-one-out cross-validation (LOOCV).Conclusion:BYOL, with the proposed custom augmentation set, can learn effective image representations without relying on a large amount of labeled data, thereby enhancing scalability since unlabeled images are easier to acquire. It surpasses BYOL with default augmentation and training under supervised learning, especially with a low-labeled data regime.