Selection of Convolutional Neural Network Model for Bladder Tumor Classification of Cystoscopy Images and Comparison...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Bladder cancer diagnosis relies on cystoscopy, with interpretation varying among clinicians.
• Deep learning offers potential for objective and accurate image analysis.

Purpose of the Study:

• To evaluate various deep learning models for bladder tumor classification using cystoscopy images.
• To compare the diagnostic performance of the best AI model against urologists and medical students.

Main Methods:

• Trained 17 convolutional neural network (CNN) models on 3,731 cystoscopy images (2,191 tumor, 1,540 normal).
• Selected EfficientNetB0 as the optimal AI model based on performance metrics.
• Compared AI model performance with diagnostic statistics from urologists and medical students using ROC curves.

Main Results:

• EfficientNetB0 achieved 81% balanced accuracy, 88% sensitivity, 74% specificity, and 92% AUC.
• The AI model outperformed medical students (69% balanced accuracy) but was less accurate than experienced urologists (91% balanced accuracy).
• EfficientNetB0 showed higher specificity (74%) compared to medical students (44%).

Conclusions:

• EfficientNetB0 is a suitable AI model for classifying bladder tumors in cystoscopy images.
• This AI technology can assist less experienced clinicians in diagnosis.
• Image-based deep learning holds potential for broader biomedical image analysis and clinical decision support.