Interpretable Model Based on CT Radiomics and Deep Learning Features for Distinguishing Typical Signs of Secondary Pulmonary Tuberculosis

Abstract

Secondary pulmonary tuberculosis presents with diverse and complex signs on CT images, which have hindered accurate diagnosis. To address this, we developed an interpretable voting model that integrates radiomics and deep learning features to classify the five most common signs of secondary tuberculosis, improving diagnostic efficiency with high accuracy. This model assists physicians in early diagnosis, lesion progression monitoring, and prognostic assessment. In this retrospective study, features of five main CT signs of secondary pulmonary tuberculosis (Tree-in-bud pattern, nodule, consolidation, thick-walled cavity, and fibrous lesion) were extracted from 350 patients CT sequences using radiomics and ResNet. 350 slices were used for training, and 150 slices from different patients were used for testing. Morphological analysis based on radiomics, SHAP analysis, Grad-CAM visualization, and statistical analysis was employed to enhance the interpretability of the model. The results indicated that the combined models statistically outperformed the individual radiomics and ResNet feature models in identifying the five main signs of secondary pulmonary tuberculosis. The AUC values (for radiomics, neural network, and combined model) on the test set were as follows: Tree-in-bud pattern: (0.795, 0.845, 0.880), Nodules: (0.830, 0.818, 0.851), Consolidation: (0.745, 0.799, 0.821), Thick-walled cavity: (0.789, 0.820, 0.814), Fibrous lesion: (0.854, 0.846, 0.934). The combined models outperformed individual radiomics and ResNet feature models in identifying the main signs of secondary pulmonary tuberculosis. The interpretability of the models was enhanced through various analysis methods. The model shows potential for improving diagnostic accuracy and supporting early diagnosis, treatment monitoring, and prognostic assessment.