Introduction: Non-muscle invasive bladder cancer is the most common clinical condition observed during the diagnosis of bladder cancer and is often a chronic condition that requires frequent cystoscopy. We developed a cystoscopy diagnostic support system using artificial intelligence (AI) to reduce tumor lesion oversight during flexible cystoscopy in clinics. In this study, we developed a deep learning-based image identification method for tumor lesions and validated the quality of the probability maps showing lesion distribution.
Methods: We developed an AI model that can calculate the probability of bladder tumor presence when using white-light (WLI) or narrow-band imaging (NBI) cystoscopy, or both. We used a dataset of 6,523 WLI and 1,636 NBI cystoscopy images (5,480 normal and 1,045 tumor images; and 1,354 normal and 282 tumor images, respectively), and divided them into training and testing datasets in an 8:2 ratio to create our AI model based on the deep learning model ResNet50 pre-trained using ImageNet. A probability map showing the distribution of the abnormal lesions in the analyzed cystoscopy images was compared with the annotation data pixel-by-pixel to measure the diagnostic accuracy of the proposed AI model. The results of AI were considered accurate if the ratio of overlap between the estimated area and the area annotated by the doctor exceeded 50%; otherwise, the results were considered inaccurate. The evaluation was performed using a five-fold cross-validation test.
Results: For the test dataset, containing 1,650 cystoscopy images (WLI 217 and NBI 64 images of bladders with tumors and WLI 1,104 and NBI 265 images of normal bladders), and at the threshold where the F1 score in the training dataset was the maximum, the proposed AI model exhibited average sensitivity, specificity, and F1 scores of 83.7%, 82.3%, and 0.637, respectively. We adapted the AI model to all frames of a clinical cystoscopy video of a case and detected the lesion sites.
Conclusions: By applying the analysis of still images to video, real-time detection of bladder tumor lesions for both WLI and NBI is possible, with AI showing the probability map of the lesion sites using the proposed method. This may improve diagnostic accuracy by creating awareness among urologists of all skill levels and by encouraging more robust observations.
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) and was supported in part by the Japanese Foundation for Research and Promotion of Endoscopy.
.jpg)
.jpg)