Distinguishing lung cancer from atypical tuberculosis: Deep transfer learning and imaging omics features

Abstract To address the diagnostic challenge posed by overlapping features, we created a deep learning (DL) model for non‐invasive differentiation of lung cancer (LC) from tuberculosis using clinical and CT data. A total of 229 patients at the Affiliated Cancer Hospital of Harbin Medical University had their clinical and CT data that were retrospectively gathered. In order to get areas of interest (ROIs), lung window CT images were manually segmented. Extracted features included clinical variables, CT semantic descriptors, radiomic profiles, and 2D/3D DL features. Logistic regression models were employed to assess the diagnostic performance of each set of features. The model incorporating 3D‐DLF extracted via a 3D‐ResNet network demonstrated the best performance. In the training cohort (sensitivity = 0.976, specificity = 0.961) and the test cohort (sensitivity = 0.789, specificity = 0.935), it obtained an Area Under the Curve (AUC) of 0.992 (95% Confidence Interval (CI): 0.984–1) and a AUC of 0.963 (95% CI: 0.929–0.998). Performance in the external validation cohort yielded an AUC of 0.709 (95% CI: 0.542–0.876; sensitivity = 0.875, specificity = 0.286). The 3D‐ResNet model outperformed those using clinical, semantic, and conventional radiomic features, highlighting DL's potential to enhance computer‐aided differentiation of LC and tuberculosis.