A Multitask Deep Learning Pipeline for Classifying Hypertrophic Cardiomyopathy and Hypertensive Heart Disease Based on Native T1 Mapping

Differentiating hypertrophic cardiomyopathy (HCM) from hypertensive heart disease (HHD) on cardiac magnetic resonance (CMR) native T1 mapping is clinically important, yet the two conditions overlap morphologically and global T1 values provide only moderate discrimination. We developed and rigorously evaluated a multitask deep learning pipeline for HCM/HHD classification. In this retrospective single-center study, 174 patients (121 HCM, 53 HHD) with 3.0-T modified Look-Locker inversion recovery (MOLLI) native T1 maps were analyzed under a two-stage protocol that strictly separated model selection from patient-level stratified fivefold cross-validation. The multitask model, comprising a shared ImageNet-pretrained ConvNeXt-base encoder, a lightweight feature pyramid segmentation head, and myocardium probability-weighted attention pooling, was compared with single-task models, a multivariable clinical logistic regression baseline, and state space (Mamba) architectures. The final model achieved an area under the receiver operating characteristic curve (AUC) of 0.941 (95% confidence interval CI 0.903-0.979), numerically higher than the best single-task model (0.902, ΔAUC + 0.039, bootstrap CI + 0.007 to + 0.074) and the clinical baseline (0.757, p < 0.0001), with consistent gains across four cross-validation seeds (mean 0.929). Its segmentation branch reached a Dice coefficient of 0.823, and all Mamba baselines were inferior (AUC 0.633-0.719). On an independent external test set of 20 patients, the model achieved AUC 0.885 (sensitivity 0.917, specificity 0.625). A multitask deep learning pipeline with segmentation-derived spatial regularization provides accurate HCM/HHD classification from native T1 maps, exceeding single-task and clinical baselines; larger multicenter validation is required.