Self-supervised learning for non-alcoholic fatty liver disease diagnosis using ultrasound imaging

Background: Non-alcoholic fatty liver disease is one of the leading causes of liver-related morbidity that can progress to severe liver damage such as liver fibrosis, cirrhosis, and liver cancer if not diagnosed and treated early. Ultrasound imaging is a non-invasive diagnostic tool that is commonly used for diagnosis. However, differences in imaging devices and the lack of labeled datasets hinder the development of generalizable machine-learning models for automated diagnosis. Objective: This thesis aims to investigate the use of Self-Supervised Learning (SSL) methods, specifically Bootstrap Your Own Latent (BYOL) and Simple Contrastive Learning of Visual Representations (SimCLR), to improve diagnostic accuracy while minimizing reliance on labeled data. Thus, reducing the time and cost required by radiologists to annotate the images. Methods: BYOL and SimCLR have been employed to learn good image representations from unlabeled images, utilizing ResNet-50 and ResNet-101 architectures to evaluate the impact of model size on classification performance. In addition, default and custom augmentation with balanced and imbalanced class distribution protocols with different batch sizes have been tested using both linear and fine-tuning evaluation with varying percentages of labeled data. Results: BYOL with ResNet-50 achieved the highest accuracy, utilizing the proposed custom augmentation set and balanced class distribution protocol, average accuracies of 91.71, 90.91, and 86.64 were achieved using 100%, 10%, and 1% of the labels with linear evaluation, respectively over three shuffled subsets. This is a statistically significant difference (P < 0.05) compared to its supervised learning counterpart with a small portion of labels. It is higher than its supervised learning counterpart by 10.47 and 16.47 using 10% and 1% of the labels, respectively. Moreover, using leave-one-out cross-validation, the same method achieved average accuracy and AUC of 97.81 and 0.971, respectively, with full label utilization. A key finding of this research is that the proposed custom augmentation set significantly improved performance, achieving the highest accuracy and AUC among all methods tested. Furthermore, BYOL shows a higher ability to handle class imbalance than SimCLR. Conclusion: BYOL, with the custom augmentation, can learn high-quality image representation without relying on labels, highlighting the potential of self-supervised learning in medical imaging applications, particularly for datasets with limited annotations.