Abstract
Breast ultrasound is a low-cost, widely accessible screening modality. However, the development of accurate deep learning classifiers for breast ultrasound is limited by the lack of expert-annotated data. Self-supervised learning (SSL) offers a pathway to high-quality representations without labels, and is widely reported to be most useful precisely when labeled data is scarce. We validate this premise on the public BUSI breast ultrasound dataset using SimCLR, assessing both linear probing and full fine-tuning on a labeled-data efficiency curve (10%–100% of available labels) against two baselines: an encoder trained from scratch (random initialization) and an ImageNet-supervised encoder (fully finetuned). Contrary to the common low-label narrative, we find that a randomly initialized ResNet-50 trained on as few as 62 images already reaches a macro AUC of 0.863, and that ImageNet-supervised fine-tuning outperforms SimCLR at every label fraction tested, including the lowest
(10%: 0.962 vs. 0.747 AUC). SimCLR fine-tuning still consistently outperforms linear probing and yields non-trivial representations from zero labels. Increasing the label budget reduces the gap to supervised pretraining, but does not close it. We argue this is a property specific to BUSI and SimCLR in this configuration: findings may not generalise to other ultrasound datasets or SSL paradigms. We complement these quantitative results with Grad-CAM analysis of attention patterns across methods, and release all code, configurations, and pretrained checkpoints to enable reproducibility.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/Deep_Brea3th_013.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=oT68VpxZPD
BibTex
@InProceedings{ZizAsm_How_MICCAISAT2026,
author = { Zizaan, Asma AND Idri, Ali},
title = { { How Far Does Contrastive Self-Supervision Get You on Breast Ultrasound? A Label-Efficiency Study on BUSI } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17256},
month = {pending},
page = {pending}
}