Guang Yang, Wentian Xu, Siyu Wang +3 more
When part of the heart muscle is damaged by a heart attack, that region stops moving the way healthy muscle does, and cardiologists reading an ultrasound look for exactly that stiffness. MCF-Net pins down where the damage is by combining image features from EchoPrime, a pretrained heart-ultrasound model, with the motion of the heart wall itself. Reading from two angles helps, since a single view can be ambiguous, especially the apical ones.
The nice part is that they keep labeling cheap. Rather than densely annotating motion, they hand-mark one template frame and let point tracking carry it across the video, then turn that into rough masks flagging which segments need a closer look. On locating infarcts at the segment level they report 72.4% F1 and 84.9% accuracy, ahead of motion-only, vision-only, and other fusion baselines.
These figures come from the abstract on their own evaluation, so treat it as a research result rather than a validated clinical tool, and see the paper for details.
Myocardial infarction (MI) remains a leading cause of mortality worldwide. Echocardiography (Echo) is a widely available modality for MI assessment, where regional wall motion abnormality is a key indicator. Prior learning based methods for myocardial motion analysis often use handcrafted descriptors or densely supervised estimation, but the need for extensive annotation limits applicability. Foundation models have recently improved vision-based Echo analysis; however, most…
When Does Muon Help Agentic Reinforcement Learning?
arXiv (cs.LG) · July 17, 2026Behaviour-Conditioned Neural Processes for Adaptive Residential Short-Term Load Forecasting
arXiv (cs.AI) · July 17, 2026An Exam for Active Observers
arXiv (cs.LG) · July 17, 2026PRISA: Proactive Infrastructure LiDAR Framework for Intersection Safety Assessment