A Lesion-Guided Deep Learning Framework for Differential Diagnosis of Pulmonary Nodules: A Pathology-Confirmed Study of Pulmonary Tuberculosis, Lung Cancer, Pulmonary Cryptococcosis, and Other Granulomatous Lung Lesions

Abstract Background Pulmonary tuberculous nodules frequently exhibit imaging features that overlap with those of lung cancer, pulmonary cryptococcosis, and other granulomatous lesions, making accurate preoperative diagnosis challenging and often resulting in unnecessary surgical intervention. We developed a lesion-guided deep learning framework to improve the differential diagnosis of surgically resected pulmonary nodules on chest CT. Methods This single-center retrospective study included 1,084 patients with surgically confirmed pulmonary nodules, comprising pulmonary tuberculosis (n = 240), pulmonary cryptococcosis (n = 298), other granulomatous inflammatory lesions (n = 252), and lung malignancies (n = 294). We developed a lesion-guided dual-branch UNet (LG-DU-Net) that jointly performed lesion segmentation and classification in an end-to-end framework. Model performance was compared with five representative deep learning models using multiple binary and multi-class classification tasks. Results LG-DU-Net consistently achieved the best diagnostic performance across all classification tasks. In benign–malignant classification, the model achieved an AUC of 0.880. The highest performance was observed in differentiating pulmonary tuberculosis from pulmonary cryptococcosis, with an AUC of 0.972 and an accuracy of 89.8%. In the three-class classification task involving pulmonary tuberculosis, pulmonary cryptococcosis, and other granulomatous inflammatory lesions, the model achieved a macro-average AUC of 0.862. Grad-CAM visualization demonstrated that model predictions were primarily driven by lesion-related imaging features. Conclusion LG-DU-Net enables accurate and interpretable differentiation of pulmonary tuberculosis, lung cancer, pulmonary cryptococcosis, and other granulomatous pulmonary nodules using chest CT images. The proposed framework has the potential to improve preoperative diagnostic accuracy, facilitate appropriate clinical management, and reduce unnecessary surgical intervention.