Dual-generative synthesis framework: enhancing polyp segmentation in colonoscopy via mask-conditional GANs

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Colorectal cancer screening relies on colonoscopy for polyp segmentation.
• Detecting small, flat polyps is challenging due to subtle features and limited training data diversity.
• Existing deep learning and generative methods have limitations in controlling structure and aligning masks with images.

Purpose of the Study:

• To present a dual generative synthesis framework for data-centric augmentation.
• To improve the segmentation performance of small and flat polyps in colonoscopy images.
• To address limitations of current generative methods regarding control, alignment, and efficiency.

Main Methods:

• A two-step framework: procedural generation of realistic polyp masks, followed by a mask-conditioned GAN for image synthesis.
• Ensures anatomically realistic and perfectly aligned mask-image pairs.
• Incorporates generated data into a U-Net segmentation model and evaluates using public datasets.

Main Results:

• Achieved a Dice score of 0.8786 and Intersection over Union (IoU) of 0.7835.
• Obtained precision of 0.8930 and recall of 0.8648.
• Demonstrated significantly higher performance compared to the baseline U-Net model.

Conclusions:

• The dual generative synthesis framework enhances segmentation robustness for small, flat polyps.
• Generates realistic and aligned training data, improving automated polyp segmentation reliability.
• Shows potential for advancing colonoscopy image analysis in cancer screening.