Motion-guided semi-supervised learning with dynamic kernels for cardiac cine segmentation

Objective: Accurate segmentation of full cardiac cine magnetic resonance (CMR) sequences is crucial for clinical diagnosis. Although deep learning models have become the standard for this task, they typically require exhaustive frame-by-frame supervision, making manual data annotation prohibitively time-consuming. Approach: To address this, we propose a novel motion-guided semi-supervised spatiotemporal framework that propagates sparse supervision from the label of the end-diastolic (ED) frame to the entire sequence. Our framework jointly optimizes motion estimation and segmentation. First, a motion decoder estimates dense displacement fields to warp the ED label, generating high-fidelity pseudo-labels for unannotated frames. Concurrently, the segmentation network employs a Temporal Kernel Segmentation (TKS) module with dynamic convolutions to adaptively capture rapid cardiac deformations, alongside a Multi-Scale Spatial Propagator (MSP) to recover fine anatomical boundaries. The motion-derived pseudo-labels directly supervise the segmentation network, ensuring temporal coherence. Main results: Extensive experiments on a multi-center cohort from three clinical centers demonstrate that our approach significantly outperforms the compared methods, achieving superior segmentation accuracy and robust generalization capabilities. Significance: The proposed framework effectively mitigates the burden of manual data annotation while providing temporally coherent full-cycle cardiac analysis, holding significant potential for efficient and reliable clinical assessment of cardiovascular diseases.