Abstract
Therapy induced cardiotoxicity is the leading nononcologi- cal cause of treatment interruption in breast cancer patients, yet pre- dicting future cardiac injury from a single pretreatment echocardiogram remains an open problem. We present EchoSense, a framework for Task 3 of the EchoRisk MICCAI 2026 Challenge that decouples feature ex- traction from temporal modelling. Per frame embeddings are extracted using a frozen domain pretrained EchoCLIP encoder, and a lightweight multiscale 1D ResNet temporal head is trained entirely from scratch on the resulting 512 dimensional sequences. We introduce preprocessing augmented training, which stacks training records derived from comple- mentary preprocessing variants to expose the temporal head to a richer distribution of input features without requiring additional patient data. Under nested five fold cross validation, where all training decisions in- cluding early stopping, hyperparameter selection, seed retention, and temperature calibration are made on inner folds and the 39 patient val- idation set is opened once for read only inference, EchoSense achieves an unbiased pooled AUROC of 0.703, substantially above the challenge baseline (test AUROC 0.541). Validation set model selection, used to guide development, yields AUROC values of 0.813 to 0.824, quantify- ing the selection optimism inherent in evaluating many configurations against a small held out set. A systematic view comparison confirms that the A4C view alone outperforms dual view averaging under the EchoCLIP feature representation. This submission is presented by Team EchoSense.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EchoRisk2026_022.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=C2ETrkl0rM
BibTex
@InProceedings{MahHas_EchoSense_MICCAISAT2026,
author = { Mahmood, Hassan AND Moniruzzaman, MD AND Hosseini, Seyed Hesamoddin AND Islam, Syed Mohammed Shamsul AND Dwivedi, Girish AND Ihdayhid, Abdul Rahman},
title = { { EchoSense: Leveraging Foundation-Model Embeddings and Multiscale Temporal Modelling for Early Cardiotoxicity Prediction } },
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}
}