In this thesis we propose a framework for the detection and classification of colorectal polyps to assist endoscopists in bowel cancer screening. Such a system will help reduce not only the miss rate of possibly malignant polyps during screening but also reduce the number of unnecessary polypectomies where the histopathologic analysis could be spared. Our polyp detection scheme is based on a cascade filter to pre-process the incoming video frames, select a group of candidate polyp regions and then proceed to algorithmically isolate the most probable polyps based on their geometry. We also tested this system on a number of endoscopic and capsule endoscopy videos collected with the help of our clinical collaborators. Furthermore, we developed and tested a classification system for distinguishing cancerous colorectal polyps from non-cancerous ones. By analyzing the surface vasculature of high magnification polyp images from two endoscopic platforms we extracted a number of features based primarily on the vessel contrast, orientation and colour. The feature space was then filtered as to leave only the most relevant subset and this was subsequently used to train our classifier. In addition, we examined the scenario of splitting up the polyp surface into patches and including only the most feature rich areas into our classifier instead of the surface as a whole. The stability of our feature space relative to patch size was also examined to ensure reliable and robust classification. In addition, we devised a scale selection strategy to minimize the effect of inconsistencies in magnification and geometric polyp size between samples. Lastly, several techniques were also employed to ensure that our results will generalise well in real world practise. We believe this to be a solid step in forming a toolbox designed to aid endoscopists not only in the detection but also in the optical biopsy of colorectal polyps during in vivo colonoscopy.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:656509 |
| Date | January 2013 |
| Creators | Hadjilucas, Lucas |
| Contributors | Bharath, Anil |
| Publisher | Imperial College London |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://hdl.handle.net/10044/1/24942 |
Page generated in 0.0215 seconds