Construction of CT-assisted diagnostic model for the activity grading of pulmonary tuberculosis based on three-dimensional convolutional neural networks
WANG Tiantian
BAO Li
LIU Yuanyuan
YAO Jianxin
LUAN Jiyue
Abstract：Objective To develop and validate a 3D convolutional neural network(3D-CNN)-based CT-assisted diagnostic model for tuberculosis(TB)activity grading,aiming to improve diagnostic efficiency and accuracy.The model incorporates Grad-CAM visualization technology to provide interpretability analysis of its decision-making process.Methods Retrospective data were collected from 300 patients who underwent non-contrast chest CT scans at Jining Public Health Medical Center from January 2020 to December 2024.According to the Diagnostic Criteria for Pulmonary Tuberculosis(WS 288-2017)and comprehensive clinical evaluation(including sputum culture,pathological results,anti-TB treatment response,and follow-up imaging dynamics),patients were stratified into three groups,with 100 cases in each group:normal lung group,active TB group,inactive TB group.The 3D-CNN model was employed to extract spatial features from CT images,with model parameters optimized through cross-validation.Model performance was systematically evaluated.The Grad-CAM algorithm generated heatmaps to identify critical regions of model attention,with results validated against clinical diagnostic standards.Results The model achieved 95.00%of classification accuracy,95.30%of sensitivity,and 95.60%of specificity on the test set.Grad-CAM visualizations demonstrated high spatial concordance between model-identified regions and expert-annotated lesion areas.Conclusion The 3D-CNN-based CT-assisted diagnostic model shows high performance in TB activity grading and may serve as an effective clinical decision-support tool.The integration of Grad-CAM enhances model transparency and credibility.
Keywords：3D convolutional neural networktuberculosisactivity gradingCT assisted diagnosis
Publication Date：2025-05-20
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：7( 620-626 )