P1.03-024 Accuracy of Combined Semantic and Computational CT Features in Predicting Non-Small Cell Lung Cancer Subtype

Abstract

With improvements in molecular treatment, it is increasingly important to differentiate non-small cell lung cancer (NSCLC) subtypes, i.e., adenocarcinoma(ADCA) from squamous cell cancer(SCCA). Many patients cannot undergo invasive biopsy procedures and so non-invasive classification methods would be helpful in their management. Most studies using CT scans for this purpose have used either semantic (visual assessment of CT images by a radiologist) or computational texture features, yielding modest accuracy. We hypothesized that combined semantic and computational assessment of CT scans would improve the accuracy of CT in NSCLC classification. 67 patients (38 ADC, 29 SCCA) underwent contrast-enhanced chest CT for lung cancer staging. Tumor volumes of interests (VOI) were drawn semi-automatically. 10 qualitative semantic and 361 computational texture features were derived from the VOIs. Univariate and multivariate logistic regression models(MLRM) were developed for combinations of semantic and texture features. Sensitivity, specificity, and area under the receiver operating characteristic (AUROC) curve were computed. 10-fold cross-validation was used to prevent overfitting. Univariate models found two semantic (air-bronchogram, shape) and five texture parameters (wavelet-transform based: GLCM_Correlation, GLRL_LGRE, GLRL_LGRE, GLRL_LGRE, and original VOI-based GLSZM_ZSN[1]) to be most predictive of tumor class (p-value <0.01). Sensitivity, specificity, and AUROC for MLRM utilizing semantic features alone was 64.2%, 73.3%, and 0.76, and that of MLRM for texture features alone was 74.6%, 72.3%, and 0.79, respectively. For combined model involving semantic and texture features (i.e., air-bronchogram and GLCM_Correlation), respective values were 81.2%, 90%, and 0.9. [1] GLCM: gray-level cooccurence matrix, GLRL_LGRE: gray-level run-length matrix-derived low gray run emphasis, GLSZM_ZSN: Gray-level size-zone matrix-derived zone-size nonuniformity. Combined semantic and computational texture assessment of lung cancer CT images is highly accurate in differentiation of SCCA and ADCA.