Abstract Body: INTRODUCTION: Accurate segmentation of ventricular structures from cardiac cine magnetic resonance imaging (MRI) is fundamental to ejection fraction and ventricular volume assessment, yet performance often degrades across scanner vendors.
HYPOTHESIS: We tested whether uncertainty-aware gating of a frozen multi-agent ensemble, coupled with gradient-free topology-guided repair, can preserve clinically meaningful volumetric accuracy under cross-vendor shift.
AIMS: To develop a vendor-robust segmentation framework that localizes unreliable predictions, suppresses weak agents, and targets topological discontinuities that most directly threaten downstream volumetric measurements.
METHODS: TopoPath integrates three frozen and architecturally diverse U-net-based agents that estimate Dirichlet uncertainty, domain shift, and topology confidence. These signals drive a mixture-of-experts gate that downweights unreliable agents during inference. A gradient-free repair module then uses artificial potential fields, A* pathfinding, and constrained smoothing to reconnect broken boundaries in uncertain regions. Models were trained on Automated Cardiac Diagnosis Challenge (ACDC) dataset (n=100) and evaluated on M&Ms dataset (n=320, four scanner vendors).
RESULTS: On ACDC, the ensemble achieved Dice values of 0.87 for the right ventricle and 0.91 for the left ventricle, with left ventricular ejection fraction mean absolute error of 1.41% and right ventricular ejection fraction mean absolute error of 3.29%. Under evaluation on M&Ms, mean Dice decreased to 0.75. However, clinically relevant volumetric performance remained robust, with left ventricular ejection fraction mean absolute error of 1.42%, left ventricular end-diastolic volume correlation of r=0.991 (p<0.001). Importantly, residual errors were concentrated in localized myocardial discontinuities and false cavities rather than global volumetric drift. This dissociation between reduced spatial overlap and retained volumetric fidelity suggests that uncertainty-aware gating can protect clinically relevant outputs.
CONCLUSIONS: This gradient-free offers a practical path toward vendor-robust cardiac MRI segmentation without inference-time weight updates.