Pericoronary Adipose Tissue Radiomics from Coronary Computed Tomography Angiography Identifies Vulnerable Plaques

Figure 1. Three-dimensional (3D) visualization of IVOCT coronary-artery segments with (A) TCFA and (B) MC. A heatmap of fibrous-cap thickness is overlayed in (A), showing the TCFA region. As described in the text, TCFA (red) was defined as a plaque with a fibrous cap < 65 µm and TCFA angle > 90° for each frame. The MC was detected as described in the text. In this instance, there were three microvessels (blue) in this plaque. The microvessels’ segments were 7.4 mm in length and their diameters were approximately 10.4 µm. Multiple radiomic features captured the extent of TCFA and microvessel presence in plaques.

Figure 2. Registration of CCTA and IVOCT images. In panel (A), 3D CCTA coronary vessel (top), the straightened CCTA (middle), and the IVOCT coronary vessel (bottom). Registration was performed using our developed software, OCTOPUS. Note that the white calcified plaques in the straightened CCTA view correspond to matching calcifications in IVOCT, demonstrating good registration. In panel (B), an IVOCT axial frame of a non-calcified lesion is shown. Panel (C) shows a registered CCTA axial frame overlayed with HU colormap segmentation of PCAT.

Figure 3. Univariate CCTA-feature analysis for predicting IVOCT-TCFA. Manhattan plot of PCAT-LOI (A) and PCAT-Vessel (B) of the mean AUCs for the identification of coronary vessels with IVOCT-TCFA. The number of radiomic features with AUC > 0.5 for identification of coronary vessels with IVOCT-TCFA was 147/293 (50.2%) for PCAT-LOI and 99/341 (29.0%) for PCAT-Vessel. Of the 147 PCAT-LOI-radiomics features with AUC > 0.5, 14 (9.5%) were shape features, 21 (14.3%) were first-order statistics, and 112 (76.2%) were texture-based features (GLCM: 18, GLDM: 21, GLRLM: 16, GLSZM: 51, NGTDM: 6). Of the 99 PCAT-Vessel-radiomics features with AUC > 0.5, eight (8.1%) were shape features, thirteen (13.1%) were first-order statistics, and seventy-eight (78.8%) were texture-based features (GLCM: 11, GLDM: 16, GLRLM: 19, GLSZM: 27, NGTDM: 5).

Figure 4. Univariate CCTA-radiomics feature analysis for predicting IVOCT-MC. Manhattan plot of PCAT-LOI (A) and PCAT-Vessel (B) of the mean AUCs for the identification of coronary vessels with IVOCT-MC. The number of radiomic features with AUC > 0.5 for identification of coronary vessels with IVOCT-MC was 195/293 (66.6%) for PCAT-LOI and 203/341 (59.5%) for PCAT-Vessel. Of 195 radiomic features with AUC > 0.5 from the PCAT-LOI, 31 (15.9%) were shape features, 27 (13.9%) were first-order statistics, and 137 (70.3%) were texture-based features (GLCM: 16, GLDM: 26, GLRLM: 19, GLSZM: 63, NGTDM: 13). Of 203 radiomic features from the PCAT-Vessel, 32 (15.8%) were shape features, 19 (9.4%) were first-order statistics, and 152 (74.9%) were texture-based features (GLCM: 24, GLDM: 28, GLRLM: 21, GLSZM: 67, NGTDM: 12).

Figure 5. Multivariate analysis of CCTA PCAT-radiomic features associated with IVOCT vulnerable-plaque characteristics on IVOCT. Diagnostic performance obtained from the three-fold cross-validation with 1000 repeats for the identification of (A) IVOCT-TCFA, (B) IVOCT-MC, and (C) IVOCT-TCFA-MC (IVOCT-Vulnerable). The ROC curves of PCAT-radiomics models for PCAT-LOI and PCAT-Vessel are plotted, and the mean AUC and SD are reported.