A CNN-based Cleanliness Evaluation for Bowel Preparation in Colonoscopy

Colonoscopy is an effective measure for detecting early colorectal lesions. To obtain high-quality pictures during colonoscopy examination, the bowel preparation, or cleansing, is considered as the first and critical step. To date, the evaluation of bowel preparation quality is performed by colonoscopists in most of gastroenterology departments. This manual labeling is time-consuming, uneconomical and sometimes unreliable subject to expertise of physicians. In this work, we propose an automatic evaluation model that labels the cleanliness for bowel preparation in accordance with the Boston Bowel Preparation Scale (BBPS). A compact convolutional neural network (CNN) with 2 Densenet layers which have feature reuse mechanism embedded before the softmax classifier is utilized. The proposed model achieves promising evaluation performance on our dataset as well as on a public dataset named Nerthus where is better than the existing research results. The average correct evaluation ratio is up to 89.58% in 20 individual cases (573 images in total), where each case includes 16 images at least. The method proposed in this paper is considered to be effective and feasible by professional colonoscopists and can be used for clinical auxiliary diagnosis.