Abstract

Cardio-oncology surveillance requires estimating left ventricular ejection fraction (LVEF), classifying subclinical dysfunction, and predicting cardiotoxicity from routine echocardiograms; each task must generalise from small, imbalanced cohorts. Existing echo foundation models plateau on the LVEF task because their frozen features encode anatomical view and cardiac motion, not LVEF magnitude.

We introduce two independent interventions that carve an LVEF axis into the representation: a self-supervised SimSiam-style regulariser $\mathcal{L}_{\mathrm{cons}}$ applied along the cardiac cycle, which requires no annotation; and dense supervision from expert-drawn LV cavity masks. Both routes reach the same LVEF axis through visibly different feature geometries, so their per-exam errors decorrelate.

A four-arm ensemble combining these two mechanisms reaches MAE = 3.90 on T1 (95% CI [3.39, 4.43]), AUROC 0.792 on T2 via cross-task LVEF reuse, and 0.7158 on T3, end-to-end from raw DICOMs. T1 and T2 blend weights are cross-validated on val. Against the official baseline of 5.02 / 0.755 / 0.646 this is $-1.12 / +0.037 / +0.070$.

Because the two interventions play different roles depending on annotation availability, our central finding transfers: cohorts without expert masks should adopt $\mathcal{L}_{\mathrm{cons}}$ as the primary axis-carving intervention; cohorts with expert masks should keep it as an ensemble decorrelation regulariser. A survey of negative results covering four echo foundation models and six adaptation strategies frames why the two interventions are complementary rather than redundant.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EchoRisk2026_007.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=72VUTyMGhr

BibTex

@InProceedings{PhePhu_Orthogonalising_MICCAISAT2026,
        author = { Phenhiran, Phuree AND Samansee, Askas AND Palawan, Chamil AND Suntudkan, Supakorn},
        title = { { Orthogonalising LVEF: Cyclic Consistency and Expert Cavity Supervision } },
        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