CardioSynergyNet: A Closed-Loop Multi-Task Deep Learning Architecture for Cardiac Segmentation and Biomarkers with...

Cardiac MRI analysis typically treats segmentation, biomarker estimation, and disease classification as separate problems, even though clinical biomarkers are themselves derived from segmentation masks and diagnosis depends on both. This paper investigates whether a single jointly trained network can perform all three tasks without sacrificing accuracy on any of them.

I propose CardioSynergyNet, which is a multi-task network with a weight-shared encoder for end-diastolic (ED) and end-systolic (ES) frames. Three modules link the tasks: Cross-Phase Deformation Attention (CPDA) models ED-ES deformation, a Differentiable Biomarker Extraction Layer (DBEL) computes eight clinical biomarkers directly from the soft segmentation mask, and Class-Conditional Prototype Feedback (CCPF) conditions segmentation on the predicted disease class. The model was trained and tested on 150 patients (30 per class across five diagnostic groups), who were split patient-wise into 110/20/20 train/validation/test, yielding 1489 paired ED/ES slices.

On the test set, the model achieved a mean ED Dice score of 0.905 (computed across all four classes, including background), with the right ventricle being the hardest structure, particularly at ES. Biomarker regression was accurate for area- and mass-based quantities (R > 0.92) but weaker for ratio-based biomarkers such as ejection fraction (R = 0.66). Disease classification reached 76.65% accuracy with most confusions occurring between clinically similar disease pairs.

The joint training of segmentation, biomarker extraction, and classification achieves performance competitive with task-specific models while keeping outputs across tasks consistent with one another, supporting cross-task feedback as a viable direction for integrated cardiac MRI analysis.