Introduction: While-light cystoscopy (WLC) has inadequate performance for detection of flat lesions such as carcinoma in situ. Blue-light cystoscopy (BLC) has improved flat-lesion detection, however adoption is modest due to equipment cost, drug instillation, and user experience. Here we develop a non-invasive, deep learning-based algorithm to augment WLC flat lesion detection by using blue-light informed transfer learning. Methods: CystoFlatNet was developed based on CystoNet, a deep learning platform for WLC bladder cancer detection. CystoNet was enhanced to augment flat lesion detection in two stages. First, a WLC image was input into a BLC domain using domain adaption of CycleGAN. Second, a region of interest (ROI) detector was used to segment the regions within the transferred images occupied by a flat lesion. Transfer learning was applied to embed features learned from both WLC and BLC into the model weights. Algorithm outputs included an outline deployed in WLC and a “mock” BLC output. For algorithm development, images of WLC and BLC matches from 50 subjects with flat lesions were selected and annotated. A training set consisting of 40 subjects and a test set of 10 subjects were used. Training was performed on WLC and BLC images. WLC was the only input used for validation. Per-lesion sensitivity was determined and performance compared to CystoNet. Results: CystoFlatNet detected flat lesions on WLC with a 90% per-lesion sensitivity (9/10). In comparison, the previously trained CystoNet had a 30% per-lesion sensitivity (3/10). Large accuracy gains were obtained in ROI overlap and visual guidance in the blue-light domain (Figure 1). Conclusions: We have developed a deep-learning algorithm that shows promise in WLC flat lesion detection. CystoFlatNet may aid in diagnostic decision-making and could serve as a non-invasive, accessible alternative to available adjunct technologies. Validation in a large prospective cohort is needed to further evaluate CystoFlatNet performance. SOURCE OF Funding: Department of Veterans Affairs Merit Review Award I01 BX005598, Urology Care Foundation Research Scholar Award
