Cystoscopic Imaging for Bladder Cancer Detection Based on Stepwise Organic Transfer Learning with a Pretrained...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Cystoscopy is crucial for diagnosing and monitoring nonmuscle-invasive bladder cancer.
• Limited cystoscopic images hinder the development of AI for bladder tumor detection.
• Transfer learning offers a potential solution to data scarcity in AI model training.

Purpose of the Study:

• To evaluate the effectiveness of stepwise transfer learning using general and gastroscopic images for enhancing AI-driven bladder tumor detection.
• To compare the diagnostic performance of the AI model with that of medical professionals.

Main Methods:

• A convolutional neural network was trained sequentially with general images, gastroscopic images, and finally, cystoscopic images of bladder tissue and tumors.
• The model's diagnostic accuracy was assessed using receiver operating characteristic curve analysis.
• Performance was benchmarked against medical students and urologists of varying experience levels.

Main Results:

• The AI model developed with stepwise transfer learning achieved 95.4% sensitivity and 97.6% specificity.
• Its diagnostic performance surpassed other models and was comparable to expert urologists.
• The model demonstrated superior accuracy in detecting tumors occupying more than 10% of the image.

Conclusions:

• Stepwise transfer learning is valuable for AI applications with limited medical image datasets.
• AI, particularly deep learning, can be a powerful tool for accurate bladder tumor detection.
• Pretraining AI models with diverse image datasets improves diagnostic capabilities in medical settings.