ncLVMass-Net: dedicated AI for left ventricular myocardium segmentation and hypertrophy detection from non-contrast cardiac CT

Background On non-contrast cardiac CT (NCCT) images, the left ventricular (LV) myocardium and blood pool share nearly identical attenuation, rendering endocardial boundary delineation infeasible. Publicly available AI models trained on contrast-enhanced CT (CECT) have been applied to NCCT for LV mass estimation, but their segmentation accuracy on this fundamentally different modality has not been rigorously validated. Objectives To develop and comprehensively evaluate ncLVMass-Net, a dedicated AI model for LV myocardium segmentation in NCCT, with respect to segmentation accuracy, LV mass estimation, subgroup consistency, and LV hypertrophy (LVH) detection. Methods CECT-based myocardial annotations from 1,529 patients were transferred to paired NCCT images through rigorous three-dimensional non-rigid registration with systematic quality verification, creating a large-scale training dataset with pixel-level ground truth. ncLVMass-Net was developed using the nnU-Net framework and evaluated against TotalSegmentator, MOOSE 3.0, and Mask R-CNN on internal (N=190) and external (N=308, three hospitals) testing sets. Performance was assessed for both thick- and thin-slice NCCT images. Results ncLVMass-Net achieved a Dice similarity coefficient (DSC) of up to 0.88 in both internal and external testing, substantially outperforming TotalSegmentator (DSC: 0.78–0.83), MOOSE (DSC: 0.67–0.78), and Mask R-CNN (DSC: 0.74–0.77). For LV mass estimation, ncLVMass-Net achieved the lowest mean absolute error (internal: 4%; external: 6%), while other models demonstrated persistent over- or underestimation with errors exceeding 12%. ncLVMass-Net reduced bias magnitude by up to 91.5% and mean absolute error by up to 78.5% relative to CECT-trained models applied directly to NCCT. Unlike TotalSegmentator and MOOSE, which exhibited progressive underestimation with increasing LV mass, ncLVMass-Net demonstrated no mass-dependent bias and maintained consistent performance across subgroups dichotomized by LV size and myocardial volume (variation < 2.3% in mean absolute error). ncLVMass-Net achieved excellent agreement with ground truth (ICC > 0.9) and yielded a superior LVH detection F1-score of up to 0.9. Conclusions A dedicated AI model can achieve accurate myocardial segmentation and quantification from NCCT despite the invisible endocardial boundary. ncLVMass-Net substantially outperforms publicly available models in segmentation, LV mass quantification, and LVH detection, enabling reliable LV mass quantification and LVH detection from routine NCCT.