Deep Learning Myocardial Segmentation in 3D Whole-Heart Joint T1/T2 Mapping: Comparison of nnU-Net and MA-SAM

This study compares two advanced deep learning models, nnU-Net and MA-SAM, for automatic segmentation of the left ventricle (LV) myocardium using 3D whole-heart T1 and T2 mapping, with the goal of evaluating their performance in terms of segmentation metrics and computational efficiency. The dataset consisted of 3D whole-heart joint T1/T2 maps from 55 subjects (15 healthy and 40 with suspected cardiovascular disease), with manual segmentations performed by an experienced clinical reader. For model development, 45 subjects were used for training, while five subjects were used for validation in each run. A cross-validation strategy with 10 independent training runs was employed, such that all subjects were used for validation across different folds. An independent hold-out test set of five subjects, which were never used during training or validation, was reserved for final testing. The fourth interleaved volume, which provided the highest myocardial-blood contrast, was used in both networks. Segmentation performance was assessed using Dice similarity score, intersection over union (IoU), and Hausdorff distance at 95%. T1 and T2 maps were generated by matching water images to a simulated dictionary and compared to manual segmentations. Both models achieved comparable segmentation performance, with nnU-Net obtaining a Dice score (DSC) of 0.91 compared to 0.90 for MA-SAM, and IoU values of 0.83 and 0.82, respectively. Regarding boundary accuracy, nnU-Net achieved a lower Hausdorff distance (HD95) of 1.64 mm, compared to 1.77 mm for MA-SAM. In addition, nnU-Net demonstrated markedly greater computational efficiency, requiring significantly less training time (5 h vs. 9 h for MA-SAM) and shorter inference time (3.67 s vs. 76.67 s), underscoring its suitability for clinical applications.