Colon Wall Segmentation

Ulcerative colitis is a chronic inflammatory bowel disease characterized by the development of ulcers in the colonic mucosa. It causes debilitating symptoms and is characterized by idiopathic periods of symptom exacerbation, followed by periodic intervals of symptom improvement or remission [5]. Therefore, therapy and medication are fluid and must always be tailored to the current disease activity which is currently assessed by endoscopic procedures such as colonoscopy or sigmoidoscopy. Recently, ultrasound of the colon wall has emerged as a reliable alternative for monitoring disease activity of ulcerating colitis. However, locating and segmenting the colon wall in ultrasound is particularly demanding and transabdominal ultrasound of the colon wall requires specialized training and expertise, which thus far is hindering wide-spread application.

Research Project

To this end, a computer-aided colon wall localization and segmentation tool has the potential to translate the advantages of transabdominal ultrasound to clinical care routines for ulcerating colitis, eliminating the need for specialized training and years of examination experience. It would allow untrained examiners to closely monitor disease activity, providing important information to guide therapeutic decisions and improve patient wellbeing

However, segmenting the colon wall in ultrasound data is a particularly difficult learning task. The colon wall is not as easily accessible as other organs because it is located deep within the abdomen and surrounded by other organs and tissue. Moreover, it is relatively thin, structurally similar to surrounding tissue and may additionally be filled with gas and feces, which can obstruct the view of the wall. In addition, an ultrasound image may depict more than one part of the colon wall, which makes it a multi-part segmentation problem and adds additional complexity.

To advance research on colon wall segmentation for transabdominal ultrasound, we collect novel dataset closely tied to the typical clinical care routine. We examine image quality and establish inter-rater variability between 4 highly specialized medical experts. In addition, we train and test several deep learning architectures for this difficult segmentation task. As a first reference performance, we see that deep learning model achieve comparable performances as human raters, while both deep learning models and humans are highly dependent on good quality images.