Development and validation of a diagnostic radiomics model for the differential diagnosis of coexistent tuberculosis and lung cancer versus isolated lung cancer.

Background: The coexistence of pulmonary tuberculosis with lung cancer (CTBLC) and isolated lung cancer (ILC) often show similar appearances on computed tomography (CT) images, making accurate diagnosis difficult and increasing the risk of misdiagnosis. Despite the need for different treatment strategies for CTBLC and ILC, research on predicting CTBLC using machine learning remains scarce. Radiomics-based machine learning algorithms have demonstrated the potential to improve diagnostic accuracy by identifying complex imaging patterns through high-dimensional features such as shape, size, intensity, and subtle textures beyond human perception. This study aims to develop and validate a non-invasive combined model, integrating radiomics signatures with clinicoradiological features, to differentiate CTBLC from ILC using CT imaging. Methods: A retrospective study was conducted on 171 patients (71 with CTBLC and 100 with ILC). The diagnosis of ILC or CTBLC was confirmed using multiple diagnostic techniques including CT imaging, sputum smear analysis, blood tests, and pathological examination following biopsy or surgical procedures. Adenocarcinoma, squamous cell carcinoma, and small cell lung cancer types were included for ILC. CTBLC lesions were identified in the same lobe and segment as cancer. Two experienced radiologists manually segmented the regions of interest (ROI) on CT images. Three-dimensional (3D) radiomics features were extracted, out of which six important features were selected for the radiomics model. We evaluated the performance of combined model with six radiomics signatures and seven clinicoradiological features using accuracy (ACC), area under the curve (AUC), among others. We also compared its effectiveness to that of clinicoradiological model as well as radiomics model. Results: The combined model showed better efficacy (AUC =0.901, ACC =0.865, sensitivity =0.900, and specificity =0.844) than the radiomics model (AUC =0.892, ACC =0.827) and significantly outperformed the clinicoradiological model (AUC =0.710, ACC =0.596). Among the clinicoradiological features, pleural thickening, lymph node enlargement, and bronchial stenosis showed significant differences between CTBLC and ILC groups (P<0.05), which contributed to the model's performance. A radiomics nomogram integrating the Rad-score with clinicoradiological features was also developed, enhancing the radiomics model's utility for the clinical prediction of CTBLC. Conclusions: The proposed CT-based combined model effectively predicts CTBLC and distinguishes it from ILC, proving a non-invasive, efficient auxiliary tool for clinical diagnosis of CTBLC.