How Far Does Contrastive Self-Supervision Get You on Breast Ultrasound? A Label-Efficiency Study on BUSI

Breast ultrasound is a low-cost, widely accessible screening modality, but the development of accurate deep learning classifiers is constrained by the scarcity of expert-annotated data. Self-supervised learning (SSL) is therefore widely reported to be most useful precisely when labeled data is scarce. However, this premise is rarely tested with a complete comparison against both training from scratch and standard supervised transfer learning across a full labeled-data efficiency curve on the same small, domain-specific dataset. We validate this premise on the public Breast Ultrasound Images (BUSI) dataset, comprising 780 images across benign, malignant, and normal classes. We pretrain a ResNet-50 encoder using SimCLR on 624 unlabeled training images, with augmentations adapted to ultrasound, and evaluate both linear probing and full fine-tuning at 10%, 25%, 50%, and 100% of available labels. We compare against an architecturally matched encoder trained from random initialization and an ImageNet-supervised encoder, using the same stratified five-fold evaluation protocol. Performance is measured by AUC, accuracy, sensitivity, and specificity. We complement the quantitative comparison with Grad-CAM analysis to examine whether performance differences correspond to meaningful differences in model attention. Contrary to the common low-label narrative, random initialization already reaches a macro AUC of 0.863 with approximately 62 labeled training images, exceeding SimCLR fine-tuning at the same label fraction (0.747). ImageNet- supervised fine-tuning performs best at every fraction tested and reaches 0.962 AUC at 10% labels, essentially matching its full-label performance of 0.958. At 100% labels, SimCLR fine-tuning reaches 0.880 AUC, only marginally above random initialization (0.876), while consistently outperforming its linear-probe variant (0.803). Increasing the label budget therefore improves SimCLR, but does not close the gap to supervised pretraining. In our comparison, SimCLR is the only method with a meaningfully rising label-efficiency curve, making it the most, rather than the least, label-hungry method. The attention analysis adds an important qualification. Random-initialized and SimCLR models produce broad, similar activation patterns that overlap lesions but often extend into surrounding tissue. The ImageNet-supervised model frequently concentrates attention near image borders, where scanner markers and acquisition artifacts may occur. A small annotation-masking check suggests that visible overlays are not the dominant driver of its predictions, but does not exclude broader shortcut learning. The unexpectedly strong low-label performance of random initialization may likewise reflect exploitable class-correlated signal or dataset-level confounds. These findings complicate, rather than reject, the value of SSL in medical imaging. Prior studies commonly pretrain on datasets far larger than BUSI; local contrastive pretraining on a few hundred ultrasound images may be insufficient to outperform a rich natural-image prior. We argue that SSL benefits are conditional on pretraining scale, dataset structure, and the features encouraged by the pretext task. The conclusions are specific to BUSI and SimCLR and should not be extrapolated to other ultrasound datasets or SSL paradigms without further validation.