Abstract

Automated left-ventricular ejection fraction (LVEF) estimation from two-dimensional echocardiography is central to cardiotoxicity surveillance. Yet real clinical datasets expose challenges that benchmarks rarely capture: small cohorts, missing apical views, coarsely quantised labels, and severe imbalance toward normal function. We present two models on an accuracy–compute frontier: EchoMoVE, a lightweight multi-view mixture-of-experts regressor (32\,M params) we design for deployment, and a fully fine-tuned, seed-ensembled adaptation of the EchoJEPA foundation model for maximum accuracy. On EchoRisk-MICCAI 2026 Task~1 (237 patients, 630 exams), EchoMoVE attains a mean absolute error (MAE) of 4.67 on the official validation split at 8\,ms/exam and 0.6\,GB. Fine-tuned EchoJEPA reaches 4.56, the only model to significantly beat the baseline but at 5.5$\times$ the cost. A systematic study, selecting models by patient-level cross-validation, shows why no single model does better. More than twenty configurations architectures, foundation-model transfer, and imbalanced-regression losses all cluster at an irreducible label-noise floor set by $\pm$2–3-point rater noise, none separating from a well-regularised baseline beyond fold variance. Variance reduction seed ensembling is the only significant lever. Re-scoring on the clinically relevant axis, tail-rich CAMUS pretraining raises low-LVEF detection AUC (0.61$\rightarrow$0.71) while MAE stays flat, and conformal intervals expose systematic under-coverage of the low-LVEF minority. Code will be released upon acceptance.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EchoRisk2026_004.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/EchoRisk2026_004_supp.pdf

Link to Open Review

Open Review Page: https://openreview.net/forum?id=eghMdLbeDO

BibTex

@InProceedings{KolGov_EchoMoVE_MICCAISAT2026,
        author = { Kolli, Govinda AND Dukre, Adinath Madhavrao AND Razzak, Imran},
        title = { { EchoMoVE: Multi-View Mixture-of-Experts Regression for Echocardiographic LVEF Estimation under Small-Cohort and Label-Imbalance Constraints } },
        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