Deep Learning Based Detection Of Urethral Stricture: Segmentation & Classification

PurposeThe retrograde urethrogram (RUG) has been a key diagnostic tool for over a century, remaining essential despite the availability of other imaging techniques for screening, diagnosis and follow up of Urethral strictures. However, interpretation of RUG images has to be done manually and needs experience on the part of the treating urologist, which calls for a common understanding of RUGs and presents a chance to improve stricture management in a practical way. Artificial intelligence (AI) algorithms present a novel way to prevent human discrepancy while concomitantly improving the accuracy of stricture identification and classification. MethodsO_ST_ABSDatasetC_ST_ABSWe have used a balanced dataset which includes RUGs of 168 strictured cases and 178 non-strictured(healthy) cases. Task#1The primary requirement is to identify the Urethral region in any clinically obtained RUGs and detect the presence of stricture in it. We successfully deployed a Segmentation and Classification model to categorize the whole dataset as strictured or non-strictured RUGs. Task#2On obtaining superior accuracy, we effectively went on to identify the type of stricture based on their location, which is of clinical importance. ResultsWith the above-mentioned available RUG dataset from 346 cases, we could train our Deep learning model and achieve a significant accuracy of 91.53% in detection and categorizing the type of stricture. At the end, a 10-fold cross-validation yielded an accuracy of about 86.66%. ConclusionOur attempts have successfully validated that using Deep learning (DL) tools, one could readily (i) Detect the presence of stricture in a given RUG and (ii) ultimately locate and classify these strictures effectively. Thus, these Deep learning tools could be of great clinical assistance for Urinary stricture related disease management.