S1425 Enhancing Gastrointestinal Bleeding Detection in Wireless Capsule Endoscopy Using Convolutional Autoencoders

Introduction: Gastrointestinal bleeding is life-threatening and demands prompt diagnosis. Wireless capsule endoscopy (WCE) offers non-invasive visualization of the entire GI tract but generates vast image volumes that challenge manual review and risk delayed or missed lesions. Machine learning—specifically autoencoders—can learn normal tissue patterns and flag deviations, making them suitable for automated anomaly detection. This study develops and evaluates a convolutional autoencoder to detect GI bleeding in WCE images, emphasizing sensitivity, specificity, and interpretability through anomaly mapping. Methods: We constructed a convolutional autoencoder with an encoder that compresses input images into a low-dimensional latent space and a decoder that reconstructs them. The model was trained on 113 normal WCE images to capture healthy tissue appearance. During validation—using both normal and lesion images—we determined a reconstruction-error threshold to distinguish normal from anomalous images. In testing, any image whose reconstruction error exceeded this threshold was flagged as containing a potential bleeding site or lesion. To enhance interpretability, we generated anomaly maps by computing pixel-wise differences between the original and reconstructed images, applied a binary mask to highlight regions of high error, and overlaid these as heatmaps onto the originals for clear localization. Results: On the test set, the autoencoder achieved a precision of 0.83, indicating few false positives, and a perfect recall of 1.00, ensuring no bleeding or lesion images were missed. The F1-score was 0.90, reflecting a strong balance between precision and recall, and the area under the ROC curve was 0.72, denoting fair discrimination between normal and abnormal images. The anomaly maps accurately highlighted bleeding and lesion regions, providing intuitive visual evidence that improved lesion localization and supported clinical interpretation. Conclusion: This study demonstrates that a convolutional autoencoder can reliably detect GI bleeding and lesions in WCE images with high sensitivity and precision. The use of anomaly maps enhances interpretability by pinpointing areas of concern, suggesting that the model could function as an effective clinical decision-support tool. Future work should test the model on larger, more diverse datasets and refine thresholding methods to further optimize performance and ease integration into clinical workflows.