A novel convolution-transformer model for differentiating hypertrophic cardiomyopathy from Phenocopies using cardiac...

Hypertrophic Cardiomyopathy (HCM) is the most prevalent inherited cardiomyopathy. Its left ventricular hypertrophy (LVH) phenotype exhibits substantial radiological overlap with HCM phenocopies such as Hypertensive Heart Disease (HHD) and Cardiac Amyloidosis (CA), posing significant challenges to clinical differential diagnosis. Differentiating HCM from its phenocopies and normal hearts on cardiac MRI is an important but difficult task. This study leveraged a cardiovascular magnetic resonance foundation model to develop a deep learning approach that enables accurate and automated differentiation among HCM, normal controls, and HCM phenocopies.

We integrated 463 cases from a multi-center dataset comprising two public benchmarks, ACDC and M&Ms2, along with a clinical cohort from a tertiary medical center. The final cohort included 207 HCM cases, 166 normal controls, and 90 cases of HCM phenocopies (including HHD, CA, and aortic stenosis, among other etiologies). To address the challenge of differential diagnosis in HCM, a tailored network was developed based on the Convolution-Transformer hybrid backbone of the CineMA cardiac MRI foundation model. The architecture employs a dual-head multi-task learning design: a primary classification head performs three-class diagnosis (HCM vs. Normal vs. Other), while an auxiliary classification head targets fine-grained discrimination between HCM and non-HCM myocardial hypertrophy. A Boundary Prior Attention Module (BPAM) was further integrated into the convolutional encoder, utilizing left ventricular myocardial segmentation masks as anatomical priors to guide spatially selective feature modulation. A three-stage progressive transfer learning strategy was implemented, transitioning from frozen backbone to partial unfreezing and finally to global optimization, with class-weighted cross-entropy and Focal Loss to address class imbalance.

The proposed method achieved an accuracy of 85.6% ± 1.7%, a macro-averaged F1-score of 0.840 ± 0.019, and an ROC AUC of 0.954 ± 0.008 under five-fold cross-validation, significantly outperforming the randomly initialized baseline (F1-score 0.704 ± 0.021, AUC 0.853 ± 0.014). HCM detection achieved a recall of 85.5% ± 2.4%, indicating a low rate of missed diagnoses in this high-risk population. Ablation experiments demonstrated that pretrained weights, the dual-head architecture, the progressive training strategy, and BPAM contributed incremental AUC gains of +6.5%, +1.2%, +1.1%, and + 1.3%, respectively.

The CineMA-based framework incorporating dual-head multi-task learning, three-stage progressive transfer learning, and anatomy-guided boundary prior attention achieves robust and accurate differentiation among HCM, normal controls, and HCM phenocopies. The approach demonstrates promising potential for AI-assisted HCM screening and clinical decision support across multi-center, multi-device settings.