Well trained, experienced gastroenterologists in academic and high volume settings can reliably recognize 97% of pathologies in Capsule Endoscopy (CE) video. However, community physicians and infrequent users may miss up to 20%. The end goal of our proposed new line of research is to develop clinical software that provides automatic decision support to physicians who are trying to declare that a patient is pathology free or has a certain disease process. The risk for the physician - and their patients - is that of a less than optimal clinical outcome due to: 1) missing a lesion/pathology in the video and putting the patient at risk of developing a more serious condition over time, or 2) mistakenly """"""""identifying"""""""" a pathology that is not present and thus subjecting the patient to unnecessary further diagnostic or surgical procedures.
The research aims i n this proposal will enable Ikona to create a pathology prioritization image processing module. Implementing modern machine learning techniques such as Support Vector Machines (SVM) and Adaboost methodologies together with proprietary image feature analysis, this technology will assign a probability metric to every frame in the image sequence for specific pathology (lesions, ulcers, bleeding, etc) and the major landmarks in the GI tract (ileo-cecal valve, pyloric valve etc.). Filtering and sorting endoscopy image data will be done such that the images with the highest probability of containing pathology will be presented to the reviewer first. This pathology prioritized sequencing is not intended to replace the clinician in the workflow, but rather to allow the clinician to focus more time on frames with a higher potential of containing pathology. Often times, clinically significant pathology may only be present in a single frame. A single """"""""pathological"""""""" frame in the middle of a 50,000 frame sequence can easily be overlooked by a novice reviewer or a reviewer whose attention is temporarily distracted. With our proposed pathology prioritization, that single pathological frame will be identified and sorted near the beginning of the image sequence thus greatly increasing the likelihood of detection by the reviewer. Specifically for Phase I, we plan to investigate and develop different algorithms for classifying image frames and recognizing pathological and normal frames, and, algorithms for ranking frames by severity of pathology. Following the implementation of a working prototype, we will further test the clinical utility of these algorithms with human clinical capsule endoscopy videos.
Capsule Endoscopy (CE) is widely used for assessing the small intestine in obscure gastrointestinal bleeding. Experienced gastroenterologists miss 2-3% of pathologies in part due to fatigue from reviewing 50,000 frames per CE video. Less experienced reviewers miss up to 20%. We propose to reduce the risk of false negatives by developing clinical image processing software to automatically re-order the CE video frames, ranking them by the probability they contain pathology.