J Endourol. 2020 Nov 4. doi: 10.1089/end.2020.0919. Online ahead of print.
BACKGROUND: Non-muscle-invasive bladder cancer is diagnosed, treated, and monitored using cystoscopy. Artificial intelligence is increasingly used to augment tumor detection, but its performance is hindered by the limited availability of cystoscopic images required to form a large training dataset. This study aimed to determine whether stepwise transfer learning with general images followed by gastroscopic images can improve the accuracy of bladder tumor detection on cystoscopic imaging.
MATERIALS AND METHODS: We trained a convolutional neural network with 1.2 million general images, followed by 8,728 gastroscopic images. In the final step of the transfer learning process, the model was additionally trained with 2,102 cystoscopic images of normal bladder tissue and bladder tumors collected at the University of Tsukuba Hospital. The diagnostic accuracy was evaluated using a receiver operating characteristic curve. The diagnostic performance of the models trained with cystoscopic images with or without stepwise organic transfer learning was compared with that of medical students and urologists with varying levels of experience.
RESULTS: The model developed by stepwise organic transfer learning had 95.4% sensitivity and 97.6% specificity. This performance was better than that of the other models and comparable with that of expert urologists. Notably, it showed superior diagnostic accuracy when tumors occupied >10% of the image.
CONCLUSIONS: Our findings demonstrate the value of stepwise organic transfer learning in applications with limited datasets for training and further confirm the value of artificial intelligence in medical diagnostics. Here, we applied deep learning to develop a tool to detect bladder tumors with an accuracy comparable to that of a urologist. To address the limitation that few bladder tumor images are available to train the model, we demonstrate that pre-training with general and gastroscopic
images yields superior results.