Abstract
Self-supervised learning (SSL) has emerged as a powerful paradigm for medical image representation learning, reducing dependence on expert annotated datasets. However, a controlled comparison of the most prominent SSL families (reconstructive, contrastive, and hybrid) for the chest X-ray domain remains lacking. Thus, in this work, we benchmark Masked Autoencoders (MAE), Momentum Contrast v3 (MoCo v3), and Contrastive Masked Autoencoders (CMAE) by pretraining on MIMIC-CXR-JPG and evaluating on a chest X-ray triage task. To the best of our knowledge, this is the first adaptation of CMAE to this radiological domain. We evaluate representation quality through (i) linear probing with 100% labels, (ii) fine-tuning at 1%, 10%, and 100% label fractions on CheXpert, reporting AUROC and F1macro across multiple random seeds, and (iii) out-of-distribution (OOD) transfer in a few-shot regime. Our results show that MAE achieves the most consistent AUROC performance, with the clearest advantage at 10% labels and cross-anatomy OOD transfer. MoCo v3 shows a localized advantage in F1macro after post-hoc threshold selection, under the 1% label regime while all methods converge under full supervision. OOD transfer was weak overall, with MAE leading under cross-anatomy transfer, while the contrastive branch showed potential under cross-institution shift. The findings suggest that the choice of SSL strategy should be guided by the intended deployment scenario.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MLMI_053.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=KpwcEbnwZr
BibTex
@InProceedings{MarPao_Benchmarking_MICCAISAT2026,
author = { Martínez-Arias, Paola AND García-Zapirain Soto, Begoña AND Ruiz, Ibon Oleagordia AND López-Linares Román, Karen},
title = { { Benchmarking Self-Supervised Pretraining Strategies for Chest X-Ray Triage and Out-of-Distribution Transfer } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17272},
month = {pending},
page = {pending}
}