Comparing Vision Transformers Towards Automatic Bowel Cleansing: A discussion

Vision Transformers (ViT) is useful in image processing when the number of training images gets large enough. The study proposes an evaluation of a series of ViT models in various configurations and from various vendors to assist the automatic assessment of the degree of bowel cleansing for colonoscopy. This approach is validated through a series of experiments conducted on three publicly available datasets (Nerthus, CVC-ClinicDB, and C3VD), as well as a set of our own collected colonoscopy frames. Bowel cleansing represents a critical preparatory step in colonoscopy, significantly influencing the diagnostic precision of endoscopic procedures, particularly in the context of colorectal cancer screening. Insufficient bowel preparation can compromise the quality of the examination, leading to ineffective procedures, misdiagnosis, and unnecessary expenditure of time and medical resources for both patients and healthcare providers. The evaluation of the degree of intestinal cleansing from the video colonoscopy is a tedious task intended to assess both the patient’s compliance with the medical indications for this exam and the precision of the medical examination. Toward automatic selection of video frames presenting specific features, Vision Transformers are valuable tools. We have selected and compared 20 ViTs models to find the most appropriate network for this classification task. Tests results are presented, and the practical configurations for running these complex structures are discussed.