Introduction: Cystoscopy is essential for the treatment of bladder cancer. However, its diagnostic accuracy is highly dependent on the experience and knowledge of the urologist. We are developing a real-time examination support system that can be connected to existing bladder endoscopy systems with a single cable to support the diagnosis of bladder cancer using artificial intelligence. During this study, we adapted our AI-support system trained cystoscopy images with a single manufacturer and validated its use with cystoscopes created by the other two manufacturers. Methods: The training data for our system with the ResNet50 model as the deep learning model comprised 2102 white-light still images (normal: 1671; tumors: 431) obtained using an Olympus VHA cystoscope. Based on the Olympus training data, our system was used with flexible cystoscopes manufactured by Karl Storz and Ambu. Two data sets containing four cases with the Karl Storz HD-VIEW and seven with the Ambu aScope 4 Cysto were used for verification. We compared the detection rate for each lesion of each case by converting the moving images to still images and comparing them with the annotated data (the correct data obtained by the expert urologist). Results: The moving images obtained using the cystoscopes manufactured by Karl Storz and Ambu resulted in 3,767 and 3,713 still images corresponding to 16 and 40 annotated tumors, respectively. Furthermore, the mean detection rates of the cystoscopes manufactured by Karl Storz and Ambu were 98.96% and 77.18%, respectively. Conclusions: Our system using images obtained with the Olympus cystoscope as training data also showed relatively good lesion detection rates when evaluating images obtained using cystoscopes manufactured by Karl Storz and Ambu. Additional training with small data set of target cystoscopy images could improve the accuracy of our system, thus providing diagnostic support for all cystoscopy systems and resulting in efficient observations and diagnoses of tumors. SOURCE OF Funding: This study is based on results obtained from a project, JPNP20006, commissioned by the New Energy and Industrial Technology Development Organization (NEDO) .
