You are here
Screening of Tuberculosis in a TB High-burden Large Rural Region in China with Deep Learning Multi-modality Artificial Intelligence.
A shortage of physicians to interpret radiological and pathological images from digital radiography and sputum smear in TB high-burden (HB) rural areas of China hinders the early diagnosis of TB. We deployed Deep-learning based Multi-modality AI (DMAI) in a TB HB large rural province, Qinhai, to assist physicians in detecting TB in radiological and pathological images. Our study investigates the efficacy of DMAI to assist physicians in detecting TB at multiple hospitals located in Qinhai.
Material and methods
A DMAI system was installed in a central TB hospital to automatically screen for TB in radiological (DR) and pathological images received from more than 60 local hospitals via secured internet connection. DMAI classified each DR image into high-risk, low-risk, and no-TB and automatically generated heatmaps denoting abnormality in <10 sec. Junior physicians (~12.5 yr-experience) reviewed these cases with DMAI support. Senior physicians (>25 yr-experience) then reviewed results from junior physicians and DMAI to generate final diagnosis. A MRMC study compared diagnoses from DMAI, junior physicians, and senior physicians from multiple hospitals to determine the effectiveness of DMAI.
Within 6 months, our DMAI system processed 105,558 radiographs and classified 77.7%, 6.4%, and 15.9% as no-TB, high-risk, and low-risk, respectively, with heatmaps in each abnormal image (Figures 1 and 2). For the high-prevalence group (age>50), DMAI classified 13% and 20% of the cases as high-risk and low-risk, respectively. DMAI and junior physicians agreed 96.3% on confirmed TB and 90.6% on non-TB cases. Compared to historical data, physicians using DMAI increased the sensitivity by 23% with similar specificity.
This is the first reported large-scale clinical application of DMAI for TB screening in China. DMAI’s performance is similar to junior physicians’. DMAI assisted physicians in detecting more TB cases in rural areas in a shorter time period without needing more physicians.