Automatic GI Bleeding Detection: A Comparative Analysis of Pre-Trained Deep Learning Architectures

The manual detection of gastrointestinal (GI) bleeding in capsule endoscopy video frames can be challenging. So, researchers are working on the development of an automatic GI bleeding detector that can help in speeding up the process of diagnosis and reduce the risk of false positives in capsule endoscopy readings. The present work presents three transfer learning (TL) architectures namely Inception-V3, ResN et-50, and Inception-ResNet-v2 for automatic GI bleeding detection using 2618 bleeding and non-bleeding capsule endoscopy frames. A comparative analysis of the TL architectures has been done through various evaluation metrics like accuracy, loss, Area-Under-Curve (AUC) and Fl-score, testing set analysis, and inter-pretability plots like gradient-weighted class activation mapping, Shapley additive explanations, and local interpretable model-agnostic explanations. ResN et50 performed best among the three TL architectures followed by Inception-V3 and Inception-ResNet-v2. ResNet50 achieved a validation accuracy, loss, AUC, and F1-score up to 91.35%, 26.50, 95.60%, and 91.32 % respectively on the validation data.