Speaker
Description
Rapid diagnostic test (RDT) kits are widely used for point-of-care detection of infectious diseases. However, interpreting their results can be challenging, especially for visually impaired individuals or when the test lines are faint or ambiguous. This paper proposes a multi-stage YOLO-based deep learning framework to automate and improve the accuracy of RDT result interpretation. The system employs three cascaded YOLO models. The first detects the overall test kit within the input image, the second localizes the membrane region where diagnostic indicators appear, and the third identifies the presence of red lines on the control (C) and test (T) regions to classify the result as positive, negative, or invalid. After each detection stage, the relevant region is automatically cropped based on bounding box coordinates to focus subsequent models on their specific areas of interest. We evaluate the effectiveness of this pipeline in terms of detection precision, classification accuracy, and the robustness of the cropping and labeling methodologies. Our results demonstrate that this approach provides a reliable and accessible tool for automated interpretation of RDT kits, potentially aiding users with limited vision or uncertain visual cues.
