Deep learning based infectivity evaluation by CXR in pulmonary tuberculosis

<bold>Background:</bold> Pulmonary tuberculosis(PTB) is a global health issue, challenged by the limitations of smear and culture tests. The introduction of CXR has advanced PTB management. <bold>Objectives:</bold> This study aims to employ deep learning techniques, particularly with CXR, to differentiate between infectious and non-infectious PTB cases. By improving both the accuracy and speed of PTB infectivity assessments, we seek to enhance public health strategies and enable more precise quarantine decisions. <bold>Methods:</bold> Employing DenseNet121 and visualization techniques like Grad-CAM++ and LIME for improved explainability, we analyzed 36,142 CXR images of 4,492 PTB patients from Severance Hospital, concentrating specifically on the lung regions through segmentation and cropping with TransUNet. We used data from 2004 to 2021 to build the model and data from 2022 to 2023 for internal validation. <bold>Results:</bold> In internal validation, the model reached an accuracy of 73.27%, boasting an AUROC of 0.7917 and an AUPRC of 0.7716. Visualization techniques such as GradCAM++ and LIME provided insights into the AI's decision-making process, enhancing both transparency and clinical applicability. <fig><object-id>erj;64/suppl_68/PA1659/F1</object-id><object-id>F1</object-id><object-id>F1</object-id><graphic></graphic></fig> <bold>Conclusions:</bold> This research introduces an AI tool that utilizes CXR for assessing PTB infectivity, potentially offering more accurate and faster screening than traditional smear tests and quicker outcomes than culture methods.