SELF-SUPERVISED LEARNING IN COMPUTER VISION: OVERCOMING DATA SCARCITY IN MEDICAL IMAGE ANALYSIS

Abstract

Deep learning has transformed computer vision, but its appetite for labelled data is a serious obstacle in medicine, where labels require expert time, privacy rules limit sharing, and rare conditions yield few examples. Self-supervised learning offers a way forward: a model first learns useful visual representations from unlabelled images by solving tasks that need no human annotation, and is then adapted to a diagnostic task with a small number of labels. This paper studies how self-supervised learning can be used to overcome data scarcity in medical image analysis. It reviews the main families of methods, which include pretext tasks such as predicting the rotation of an image, contrastive learning that pulls together embeddings of related images, and multimodal learning that links images with reports, and it examines how positive pairs can be constructed from medical data: augmented copies of one image, multiple views of the same patient, repeated samples over time, and paired imaging and text. It then proposes a staged framework that combines transfer from natural images, an intermediate self-supervised stage on large unlabelled medical collections, and fine-tuning on the target task. The framework was evaluated on chest radiograph classification, retinal disease grading, and lung computed tomography finding detection, using label fractions from 1% to 100%. With 1% of the labels, the staged framework with patient-aware contrastive learning reached an average area under the receiver operating curve of 0.861, against 0.742 for training from scratch and 0.791 for supervised transfer from natural images, and it matched fully supervised training with about one tenth of the labels. It also generalised better to an external hospital and was more robust to label noise. The results show that self-supervised pretraining on unlabelled medical data, designed to reflect the structure of clinical data, is a practical remedy for label scarcity.