Flat lesion detection of white light cystoscopy with deep learning

Abstract

Adequate lesion detection is critical in cystoscopy to improve bladder cancer resection and decrease recurrence. In particular, flat-appearing cancer such as carcinoma in situ is difficult to discern by standard white-light cystoscopy (WLC). The adoption of blue-light cystoscopy (BLC), an adjunct imaging technique, remains modest due to the expensive equipment required. We developed a deep-learning algorithm, CystoNet-F, for augmented detection of flat lesions on WLC. CystoNet-F was designed to augment WLC in lesion detection by incorporating domain translation of CycleGAN, transfer learning, and region of interest (ROI) detection. We constructed a development dataset of 40 patients for algorithm training and 10 patients for testing. In the training phase, features from both WLC and BLC were learned and embedded in the algorithm as the model weights of an ROI detector. Transfer learning was performed by fine-tuning CystoNet-F on BLC using the weights learned from WLC. We applied CycleGAN for domain translation between WLC and BLC. In the test phase, WLC input was first translated to the BLC domain and then served as the input of the ROI detector to finally generate a mask on the lesion area. CystoNet-F can produce flat lesion predictions close to urologist’s annotations on the validation set without paired BLC information. The proposed deep-learning algorithm may improve the diagnostic yield of standard WLC in a noninvasive and cost-effective fashion.