Deep learning-enabled detection and localization of gastrointestinal diseases using wireless-capsule endoscopic images

Purpose: Detection of gastrointestinal (GI) diseases involves several expensive, challenging, and time-consuming procedures. Deep learning techniques-based computer-aided diagnosis can significantly reduce the costs associated with examinations while enhancing the accuracy and speed of diagnosis. Methods: This study developed a 13-layered convolutional neural network (CNN) model named as GINet for diagnosing GI diseases such as angiectasia, lymphangiectasia, GI bleeding, and ulcer. The model was trained, validated, and tested on 3658 wireless capsule endoscopy images. Results: The GINet achieved an overall classification accuracy of 99% (standard error (SE): 0.7%) on the test dataset. The model achieved an overall sensitivity of 99.6 % (SE:0.5%) and specificity of 99.86%(SE:0.3%) on the test. We additionally employ GradCAM and Guided-GradCAM technique to enhance the interpretability and localization of detected lesions. Furthermore, the study effectively addressed biases that could arise from combining video frames of the same patients through domain adaptation techniques, ensuring accurate and unbiased analysis. Conclusions: The results revealed that the GINet can classify GI diseases with higher sensitivity and specificity. Furthermore, the proposed approach has the potential for first-hand mass screening efforts, where advanced disease stages can be identified. This approach additionally helps gastroenterologists make quick and accurate treatment decisions at reduced time and cost.