Abstract

We compare CNN (R(2+1)D-18) and ViT (DINOv2-ViT-S with LoRA adaptation) backbone families fine-tuned from EchoNet-Dynamic pretraining for all three EchoRisk Challenge tasks: (T1) LVEF regression, (T2) subclinical LV dysfunction detection, and (T3) cardiotoxicity risk prediction. A key contribution is iterative nnU-Net self-training on EchoRisk to generate per-frame pseudo-segmentations, used both as an auxiliary input channel for R(2+1)D and as per-frame LV area scalars for DINOv2. We systematically ablate multi-view fusion strategies, temporal window lengths, and segmentation-derived signals. Under 8-seed evaluation, a tail-aware conditional-minimum ensemble (Ens5-T) achieves the best Task 1 result, MAE=4.08$\pm$0.05. Qualitative trajectory analysis reveals systematic mean reversion as the primary failure mode; baseline, best individual model, and best ensemble each detect only two of nine qualifying EF-decline events. For Task 2, R(2+1)D A4C achieves AUC=0.801 (challenge baseline: 0.755), outperforming ViT for LV dysfunction detection. For Task 3, R(2+1)D trained on A4C achieves AUC=0.655$\pm$0.021 under 8-seed evaluation, only marginally above the challenge baseline (0.646). Team: ipag.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EchoRisk2026_006.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=z4PbqKuOdi

BibTex

@InProceedings{RicAnn_CNN_MICCAISAT2026,
        author = { Rickmann, Anne-Marie AND Xu, Xuanang AND Staib, Lawrence H. AND Duncan, James S.},
        title = { { CNN and ViT Backbone Fine-Tuning with Pseudo-Label Segmentations for the EchoRisk Challenge } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17268},
        month = {pending},
        page = {pending}
}


back to top