Distinguishing between Crohn’s disease and ulcerative colitis using deep learning models with interpretability

Abstract

AbstractCrohn’s disease and ulcerative colitis are two chronic diseases that cause inflammation in the tissues of the entire gastrointestinal tract and are described by the term inflammatory bowel disease. Gastroenterologists find it difficult to evaluate endoscopic images to recognise the characteristics of the two chronic diseases. Therefore, this work aims to build a dataset with images of Crohn’s disease and ulcerative colitis (collected from the public datasets LIMUC, HyperKvasir and CrohnIPI) and train deep learning models (five CNNs and six ViTs) to develop a tool capable of helping doctors to distinguish the type of inflammatory bowel disease. In addition, as these architectures will be too heavy to work in a hospital context, in this work, we are looking to use knowledge distillation to create lighter and simpler architectures with the same precision as the pre-trained architectures used in this study. During this process, it is important to evaluate and interpret the pre-trained architectures before the distillation process, and the architectures resulting from knowledge distillation to ensure that we can maintain performance and that the information learnt by both architectures are similar. It is concluded that is possible to reduce 25x the number of parameters while maintaining good performance and reducing the inference time by 5.32 s. Allied with this, through the interpretability of the models was concluded that both before and after the knowledge distillation are possible to identify ulcers, bleeding situations, and lesions caused by the inflammation of the disease.