Machine Learning Based on Multiparametric Features from Dual-Layer Detector Spectral CT for Identifying Ulcer-Like...

Aortic intramural hematoma (IMH) is a high-risk subtype of acute aortic syndrome (AAS). The presence of ulcer-like projection (ULP) significantly increases the risk of progression to dissection, aneurysm, or rupture. However, conventional computed tomography angiography (CTA) has limited reliability in differentiating small ULP from the surrounding high-density hematoma or coexisting atherosclerotic plaques. This study developed and internally evaluated machine learning (ML) models integrating spectral CT-derived quantitative parameters for ROI-based quantitative classification in IMH patients. Data from 95 IMH patients were retrospectively collected and divided into training and test sets at a 7:3 ratio. Quantitative features, including virtual monochromatic images (VMI), iodine concentration (IC), normalized iodine concentration (NIC), effective atomic number (Z_eff), and spectral curve slope (k), were extracted from spectral post-processing and subjected to LASSO regression for feature selection. Eight features (VMI 40 keV, Z_eff, IC, NIC, and four spectral curve slopes) were selected to construct six ML models. The random forest (RF) model showed the highest AUC point estimate in this internal test set, with an AUC of 0.889 (95% CI: 0.769-1.000), a sensitivity of 93.7%, and an accuracy of 82.8%. SHAP analysis revealed that Z_eff contributed most significantly to predictions, followed by the low-energy spectral curve slope, indicating that the model effectively captures differences in material composition and iodine attenuation characteristics. These findings demonstrate the exploratory internal feasibility of using multiparameter spectral CT-based ML models for ROI-based quantitative classification of ULP in IMH, complementing single-parameter analysis.