Attention-based Model for Automatic Assessment of Cleanliness in Capsule Endoscopy Videos

The diagnostic yield of Capsule Endoscopy (CE) depends on the quality of the video in routine clinical practice. Automatic cleanliness assessment may help in expediting the process of video analysis. This study presents an attention-based Visual Geometry Group 16 (VGG-16) model for automatic cleanliness assessment of CE frames. It also includes a comparative analysis of the proposed model with five Transfer Learning (TL) models using different batch sizes. The models were trained on patched CE frames and tested on both unseen patched and full CE frames. The proposed model achieved a recall of 97 % and 39 % for dirty and clean full test set frames. Gradient-weighted Class Activation Mapping (Grad-CAM), Shapley and Local Interpretable Model-agnostic Explanations (LIME) of the proposed model have been provided to highlight the regions of importance for the predictions made by the model. The proposed model has displayed the ability to identify clean frames which has been a challenge in existing state-of-the-art works.