Construction of a semi-automatic localization and qualitative diagnosis model for myocardial gray zone based on dual-energy coronary CT angiography
C Pan, L Zeng, Y Yang, Z Yang, X Zhou, T Jia, F Hu, L Lu, L Luo, B Chen, S Li, J ChenAbstract
Background
The myocardial gray zone (GZ)—the peri-infarct transition region between the myocardial scar (MS) and viable myocardium—serves as a critical electrophysiological substrate for re-entrant ventricular arrhythmias and sudden cardiac death. While cardiovascular magnetic resonance late gadolinium enhancement (CMR-LGE) remains the reference standard for GZ characterization, its clinical utility is constrained by long acquisition times and patient-related contraindications. Dual-energy coronary computed tomography angiography (DECCTA) is widely accessible, and its virtual monochromatic images (VMIs) provide richer diagnostic information than conventional CCTA. While radiomics has shown efficacy in identifying MS, its capability for the simultaneous identification and semi-automated localization of the GZ remains to be established.
Purpose
To develop and validate a semi-automatic diagnostic model based on DECCTA-derived VMIs for the localization and qualitative characterization of GZ, MS, and normal myocardium.
Methods
This prospective study included 40 patients with old myocardial infarction who underwent concurrent DECCTA and CMR-LGE. Using CMR-LGE as the ground truth,. segmentations of GZ, MS, and normal myocardium were performed on DECCTA VMIs at 40 keV, 50 keV, 130 keV, and the optimal contrast-to-noise ratio (CNR) keV,. A total of 1218 radiomics features were extracted. Feature selection was performed using minimum redundancy maximum relevance (MRMR) and LASSO algorithms with leave-one-out cross-validation. Three distinct machine learning architectures—random forest (RF), extreme gradient boosting (XGB), and k-nearest neighbour (KNN)—were deployed to construct the diagnostic models. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC).
Results
A total of 2777 annotated regions of interest (ROIs) were analyzed, partitioned into a training set (1016 GZ, 1016 MS, and 470 normal) and a validation cohort (112 GZ, 112 MS, and 47 normal). More than 86% of the most discriminative features were derived from transformed image domains, including first-order statistics, gray-level co-occurrence matrix (GLCM), and gray-level dependence matrix (GLDM) metrics. These features primarily represent high-order texture and spatial heterogeneity of the myocardium. The machine learning models demonstrated high discriminatory power, with mean AUCs ranging from 0.86 to 0.94 in the validation cohort. Notably, the RF model derived from 50 keV VMIs yielded superior performance, achieving an AUC of 0.94 for the differential diagnosis of GZ, MS, and normal myocardium.
Conclusion
Integrating radiomics features from DECCTA-VMIs with machine learning algorithms effectively enhances the differential diagnosis of GZ, MS, and normal myocardium, enabling the semi-automated assessment of the myocardial GZ.Figure 1Figure 2