Paper on Rapid Kit Test results prediction from smartphone is arXived!

AI-Driven Smartphone Solution for Digitizing Rapid Diagnostic Test Kits and Enhancing Accessibility for the Visually Impaired

Rapid diagnostic tests are crucial for timely disease detection and management, yet accurate interpretation of test results remains challenging. In this study, we propose a novel approach to enhance the accuracy and reliability of rapid diagnostic test result interpretation by integrating artificial intelligence (AI) algorithms, including convolutional neural networks (CNN), within a smartphone-based application. The app enables users to take pictures of their test kits, which YOLOv8 then processes to precisely crop and extract the membrane region, even if the test kit is not centered in the frame or is positioned at the very edge of the image. This capability offers greater accessibility, allowing even visually impaired individuals to capture test images without needing perfect alignment, thus promoting user independence and inclusivity. The extracted image is analyzed by an additional CNN classifier that determines if the results are positive, negative, or invalid, providing users with the results and a confidence level. Through validation experiments with commonly used rapid test kits across various diagnostic applications, our results demonstrate that the synergistic integration of AI significantly improves sensitivity and specificity in test result interpretation. This improvement can be attributed to the extraction of the membrane zones from the test kit images using the state-of-the-art YOLO algorithm. Additionally, we performed SHapley Additive exPlanations (SHAP) analysis to investigate the factors influencing the model’s decisions, identifying reasons behind both correct and incorrect classifications. By facilitating the differentiation of genuine test lines from background noise and providing valuable insights into test line intensity and uniformity, our approach offers a robust solution to challenges in rapid test interpretation.

Check out the paper: “archived version”