Using Synthetic and Pseudosynthetic Data to Enhance Polyp Detection in Future AI-Assisted Endoscopy Frameworks. Is it the Right Time?

Background/Objectives: Colorectal cancer (CRC) is a widespread malignancy that can be mitigated through early detection and removal of precancerous polyps. Artificial intelligence (AI), particularly deep learning, enhances polyp detection during colonoscopy but faces challenges due to limited medical imaging datasets. This study evaluates whether synthetic and pseudosynthetic data—augmented data generated from original datasets—can improve AI accuracy for polyp detection. Methods: We used multiple real and synthetic polyp datasets, applying data augmentation techniques to create pseudosynthetic data, and employing a modified U-Net architecture for polyp segmentation. The model was trained and evaluated across ten experiments using different combinations of real, pseudosynthetic, and synthetic data. Performance metrics included accuracy, precision, recall, Dice coefficient, Intersection over Union (IoU), and F1 score. External validation used the CVC-Colon-DB dataset with 612 image-mask pairs. Results: Combining real and pseudosynthetic data achieved the best performance, with a Dice score of 0.7638, precision of 0.7638, recall of 0.6774, and F1 score of 0.8979. Models trained solely on CycleGAN-generated data performed poorly, while diffusion-based synthetic data offered better generalization (precision: 0.7488, recall: 0.6695, F1: 0.8987). Conclusions: U-Net models trained with synthetic and pseudosynthetic data outperform those trained solely on real data, effectively addressing data scarcity, diversity, and ethical concerns. Models benefit more from pseudosynthetic data alone compared to mixed sources. Diffusion-generated synthetic data leads to better model performance than GAN-generated data. These findings confirm that synthetic and pseudosynthetic data are effective tools to improve model generalization and address ethical concerns in AI-assisted diagnostic environments.