Semi-Supervised Bladder Tissue Classification in Multi-Domain Endoscopic Images.

Area of Science:

• Medical Imaging
• Computer Vision
• Oncology

Background:

• Accurate visual classification of bladder tissue during Trans-Urethral Resection of Bladder Tumor (TURBT) is crucial for early cancer diagnosis.
• White Light Imaging (WLI) and Narrow Band Imaging (NBI) offer complementary visual information for lesion detection.
• A challenge exists in classifying tissue when annotations are limited to one domain (WLI) and datasets are unpaired.

Purpose of the Study:

• To develop a computer vision method for improved endoscopic diagnosis of bladder tumors.
• To address the challenge of multi-domain tissue classification with limited annotations.
• To enhance the accuracy of bladder cancer detection during TURBT procedures.

Main Methods:

• A semi-supervised Generative Adversarial Network (GAN) approach was proposed.
• The method utilized a teacher network trained on labeled WLI data.
• Cycle-consistency GAN for unpaired image-to-image translation and a multi-input student network were employed.

Main Results:

• The proposed method achieved an average classification accuracy of 0.90 for WLI and 0.92 for NBI.
• Precision and recall rates were also high, indicating robust performance across domains.
• Generated synthetic images were of sufficient quality to be indistinguishable from real images by specialists.

Conclusions:

• Semi-supervised GAN-based classification shows significant potential for multi-domain data with limited annotations.
• This approach can improve bladder tissue classification accuracy in endoscopic procedures.
• The study highlights the effectiveness of leveraging unpaired data for enhanced diagnostic capabilities.