Lung Ultrasound and Artificial Intelligence Technology for the Diagnosis of TB in LMICs

Project Summary Improved point-of-care tests and diagnostic algorithms for tuberculosis (TB) are urgently needed to enable more timely and accurate diagnosis. Currently, lack of diagnosis and diagnostic delays are significant contributors to increased mortality [1,2,3] in low and middle-income countries (LMICs), due to reliance on insensitive, slow and/or locally inappropriate tests and limited access to optimal diagnostic modalities. Fortunately, rapid advances in diagnostic imaging technology have produced affordable, portable, point-of-care ultrasound devices that can be transported with exceptional ease to resource-limited settings. Lung ultrasound (LUS) is now regularly used to accurately diagnose a variety of pulmonary disorders including pneumonia and pulmonary edema. Our preliminary studies have demonstrated 96% sensitivity of LUS for detecting associated sonographic abnormalities in patients with microbiologically confirmed pulmonary tuberculosis (PTB).5 However, there remain barriers to implementing LUS for TB diagnosis, including a scarcity of robust data about ideal training and scanning procedures. Our overall goal is implementation of real-time AI-facilitated LUS for timely evaluation of people with TB-suspected symptoms to triage who needs further evaluation and testing. Our preliminary data suggest LUS may be highly sensitive for the diagnosis of PTB, but no rigorous, adequately powered studies have investigated lung ultrasound findings in patients with PTB versus controls without PTB. To address these critical information gaps, we aim to: Aim 1. Develop a LUS model for PTB detection as a triage tool for PTB diagnosis which can be used in field settings in low-resource and remote areas. Hypothesis: LUS will have similar or better sensitivity for diagnosis of PTB compared to CXR with moderate (i.e., 70-80%) specificity when interpreted by trained personnel and validated by experts. Aim 2. Develop and test an artificial intelligence (AI) algorithm for detecting PTB by LUS that does not require trained personnel for use in LMICs. Hypothesis: An AI algorithm based on convolutional neural networks (CNNs) will classify LUS features indicative of PTB with high sensitivity (90)% vs. the reference standard of microbiological testing. This study will leverage resources and expertise among partners in the United States and Peru. Our multidisciplinary research team at the Universidad Peruana Cayetano Heredia (UPCH), the Peruvian NGO A.B. PRISMA, and the Johns Hopkins University have a strong track record of collaborative work in novel research projects in resource-poor settings, including TB and LUS. The use of portable AI-augmented LUS could save lives in resource-limited settings by decreasing time to case detection and treatment initiation.