Arxiv 2512.09779

Mohamed Elbayumi1,2, and Mohammed S.M. Elbaz1,2* Departments of 1Radiology and 2Biomedical Engineering, Northwestern University, Chicago, IL, USA *corresponding author email: mohammed.elbaz@northwestern.edu

Abstract— Few-shot learning (FSL) mitigates data scarcity in cardiac MRI segmentation but typically relies on semi-supervised techniques sensitive to domain shifts and validation bias, restricting zero-shot generalizability. We propose PathCo- LatticE, a fully supervised FSL framework that replaces unlabeled data with pathology-guided synthetic supervision. First, our Virtual Patient Engine models continuous latent disease trajectories from sparse clinical anchors, using generative modeling to synthesize physiologically plausible, fully labeled 3D cohorts. Second, Self-Reinforcing Interleaved Validation (SIV) provides a leakage-free protocol that evaluates models online with progressively challenging synthetic samples, eliminating the need for real validation data. Finally, a dynamic Lattice-of-Experts (LoE) organizes specialized networks within a pathology-aware topology and activates the most relevant experts per input, enabling robust zero-shot generalization to unseen data without target-domain fine-tuning.
We evaluated PathCo-LatticE in a strict out-of-distribution (OOD) setting, deriving all anchors and severity statistics from a single-source domain (ACDC) and performing zero-shot testing on the multi-center, multi-vendor M&Ms dataset. PathCo-LatticE outperforms four state-of-the-art FSL methods by 4.2–11% Dice starting from only 7 labeled anchors, approaches fully supervised performance (within 1% Dice) with only 19 labeled anchors. The method shows superior harmonization across 4 vendors & generalization to unseen pathologies. [Code will be made available]. Index Terms— Few-shot Segmentation, Domain Adaptation, Cardiac MRI Segmentation.
I. INTRODUCTION Cardiac Magnetic Resonance Imaging (MRI) is the gold standard for non-invasive assessment of cardiac structure and function [1], [2], [3]. However, its clinical utility relies critically on the accurate segmentation of the left ventricle (LV), right ventricle (RV), and myocardium [4]. These segmentations are prerequisites for quantifying ejection fraction, ventricular volumes, and myocardial mass—parameters indispensable for diagnosing cardiomyopathies and guiding therapy. While Deep Learning (DL) has automated this task [5], its deployment is constrained by a “data bottleneck”: fully supervised models require dense, pixel-level expert annotations that are scarce and expensive to acquire. Consequently, models trained on limited

This work was supported in part by National Heart, Lung, and Blood Institute (NHLBI) of the National Institutes of Health (NIH) Grant R01HL169780”.
Mohammed S. M. Elbaz and Mohamed Elbayumi are with the Department of Radiology and Biomedical Engineering at Northwestern University, Chicago, IL, USA. Mohammed S.M. Elbaz is the corresponding author: mohammed.elbaz@northwestern.edu

data often fail to generalize across diverse pathologies, vendors, and imaging protocols encountered in multi-center practice [6]. To mitigate annotation scarcity, Few-Shot Learning (FSL) has emerged as a promising paradigm. Current FSL approaches
in medical imaging [7], [8] predominantly adopt semi- supervised frameworks, such as meta-learning [9] or prototypical networks [7] augmented with unlabeled data. These methods infer class representations from a small labeled “support” set combined with a larger unlabeled cohort. While effective in controlled settings, this reliance on semi- supervision introduces critical limitations for real-world deployment. First, these methods assume access to a representative unlabeled cohort from target domain, which restricts zero-shot generalizability and introduces hidden biases toward the specific distribution of the unlabeled pool. Second, randomly sampled support sets rarely capture continuous spectrum of disease progression, resulting in “spectral gaps” where intermediate or extreme pathological states are underrepresented. Third, common practice of cross-validating on the labeled support set risks data leakage and inflates performance estimates in data-scarce regimes [10], [11]. In this work, we argue that the reliance on unstructured unlabeled data is a bottleneck, not a solution. We propose Pathology-Constrained Lattice-of-Experts (PathCo-LatticE), a framework that reformulates few-shot MRI segmentation from a semi-supervised to a fully supervised, data-centric paradigm. Our key insight is that cardiac disease progression follows continuous, physiologically measurable trajectories e.g., progressive wall thickening in Hypertrophic Cardiomyopathy (HCM) or ventricular remodeling in Dilated Cardiomyop

…(Full text truncated)…