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

Area of Science:

• Urology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate bladder lesion detection is critical for early cancer diagnosis and monitoring.
• Conventional cystoscopy visual inspection has limitations in detection rates.
• Deep learning offers potential to enhance diagnostic accuracy.

Purpose of the Study:

• To evaluate the HRNetV2 deep learning model for intelligent bladder lesion detection.
• To assess the model's performance across different image resolutions.
• To determine the clinical utility of AI in bladder lesion identification.

Main Methods:

• Utilized HRNetV2 semantic segmentation model on 102 white-light cystoscopy videos from 94 patients.
• Manually annotated suspected bladder lesions across 33,657 frames.
• Assessed diagnostic performance using sensitivity, precision, and mean Dice (mDice) score on high- and low-resolution images.

Main Results:

• Overall test set sensitivity: 91.6%, precision: 91.3%, mDice: 80.3%.
• High-resolution images achieved sensitivity of 94.8%, precision of 94.4%, and mDice of 84.7%.
• Low-resolution images showed lower performance: sensitivity 75.6%, precision 74.8%, mDice 56.6%.

Conclusions:

• HRNetV2 demonstrates excellent bladder lesion detection capabilities, especially with high-resolution images.
• The model shows significant potential to improve clinical detection accuracy for bladder tumors.
• Further validation with larger, multicenter datasets is recommended.