EPFL / Lausanne University Hospital (CHUV)
Self-supervised deep learning model for coronary artery segmentation from invasive X-ray coronary angiography (ICA), designed to reduce reliance on large annotated datasets. CM-UNet combines a Contrastive Masked Autoencoder (CMAE) with a UNet backbone: an online encoder-decoder branch reconstructs masked image patches while a momentum branch produces contrastive embeddings, jointly pretraining the network on unannotated angiography images before fine-tuning on a small labeled set. Fine-tuning with only 18 annotated images (instead of 500) led to just a 15.2% drop in Dice score, versus a 46.5% drop for baseline models trained without this self-supervised pretraining -- demonstrating strong label efficiency for coronary segmentation.
Architecture
Hybrid
Contrastive Masked Autoencoder (CMAE) combined with a UNet backbone: an online encoder-decoder reconstructs masked image patches while a momentum-updated branch generates contrastive embeddings for self-supervised pretraining, followed by supervised fine-tuning on limited annotated coronary angiography data
Framework
PyTorch
Added to catalog
2026-08-14
Unannotated X-ray coronary angiography corpus (CM-UNet)
Unannotated invasive X-ray coronary angiography images used for self-supervised contrastive-masked pretraining, plus a limited annotated set (up to 500 images, as few as 18 used in label-efficiency experiments) for fine-tuning.
Segmentation of coronary arteries from X-ray angiography images, using self-supervised pretraining plus limited-label fine-tuning