Segmentation-preserving warm-start knowledge distillation for efficient multi-task cardiac MRI analysis

Efficient cardiac magnetic resonance imaging (CMR) analysis requires a compact model that improves anatomical segmentation without sacrificing the diagnostic branch or increasing deployment cost. We propose a segmentation-preserving warm-start knowledge-distillation (KD) strategy for a multi-task student that is first trained for 30 supervised epochs and then refined for five epochs from its best checkpoint using a frozen higher-capacity teacher. The final objective retains ground-truth segmentation and classification supervision, transfers softened segmentation outputs with temperature τ=3 and weight λ_segKD=0.03, and omits classification-logit KD. All principal ACDC experiments were conducted under a fixed five-fold patient-disjoint protocol on NVIDIA Tesla V100 32-GB GPUs. The proposed student increased mean foreground Dice from 0.7033 ± 0.0261 to 0.7270 ± 0.0196; the paired mean gain was 0.0237 (95% CI [0.0090,0.0383], Cohen’s d_z=2.01, paired t-test p=0.0109). Classification changes were modest and statistically inconclusive. A controlled five-epoch supervised-continuation baseline reached Dice 0.7219, showing that additional refinement time explains part, but not all, of the observed gain. Patient-level out-of-fold aggregation across 100 ACDC patients increased mean Dice from 0.7985 to 0.8226 (stratified paired-bootstrap 95% CI for the gain [0.0140,0.0348]). Zero-shot segmentation on 150 labeled M&Ms patients further improved model-averaged mean Dice from 0.6643 to 0.6985, with a patient-level paired-bootstrap gain of 0.0342 (95% CI [0.0268,0.0419]). The distilled and supervised students retain exactly 3.532M parameters and 4.231 GMACs; controlled V100 profiling yielded effectively identical mean latency (1.751 versus 1.747 ms). These findings support warm-start structural distillation as a segmentation-oriented, multi-task-balanced refinement strategy with no architectural inference overhead.