Leveraging Deep Learning in Real-Time Intelligent Bladder Tumor Detection During Cystoscopy: A Diagnostic Study.

Abstract

Accurate detection of bladder lesions during cystoscopy is crucial for early tumor diagnosis and recurrence monitoring. However, conventional visual inspection methods have low and inconsistent detection rates. This study aimed to evaluate the effectiveness of the HRNetV2 deep learning model for intelligent bladder lesion detection, focusing on its performance at different image resolutions. We recruited 94 patients undergoing cystoscopy or transurethral resection of bladder tumor and collected 102 white-light cystoscopy videos between July 2022 and July 2023. Frames containing suspected bladder lesions were manually annotated. HRNetV2, a semantic segmentation model, was used to identify tumor-related morphological features. Images were categorized as high- or low-resolution, and the dataset was split into training and test sets (4:1 ratio). Diagnostic performance was assessed using sensitivity, precision, and mean Dice (mDice) score. A total of 33,657 frames were annotated. The overall sensitivity and precision for the test set were 91.6% and 91.3%, respectively, with an mDice score of 80.3%. In the high-resolution group, sensitivity and precision were 94.8% and 94.4%, while in the low-resolution group, they were 75.6% and 74.8%. The mDice scores for high- and low-resolution images were 84.7% and 56.6%, respectively. HRNetV2 demonstrated excellent performance, particularly with high-resolution images, offering high sensitivity and precision in bladder lesion detection. The results suggest that HRNetV2 has significant potential to improve detection accuracy in clinical practice. Future work should focus on optimizing the model further and testing it with larger, multicenter datasets.