Examining the effect of synthetic data augmentation in polyp detection and segmentation

Area of Science:

• Medical Image Analysis
• Deep Learning
• Gastrointestinal Imaging

Background:

• Deep learning for medical image analysis, particularly gastrointestinal imaging, faces challenges due to limited data, privacy issues, and insufficient pathology samples.
• Data augmentation is crucial for improving the performance and generalization of deep learning models in this field.

Purpose of the Study:

• To investigate the generation of synthetic colonoscopy images with polyps using a pix2pix model for data augmentation.
• To evaluate the utility of these synthetic samples in enhancing deep learning models for polyp segmentation and detection.

Main Methods:

• A pix2pix Generative Adversarial Network (GAN) was modified and trained to create synthetic colonoscopy images containing polyps.
• U-Net and Faster R-CNN models were trained on datasets augmented with varying quantities of synthetic and traditional samples.
• Model performance was evaluated using F1 score, intersection over union, precision, recall, and generalization ability on unseen data.

Main Results:

• Increasing the number of synthetic samples improved the F1 score and intersection over union for the U-Net segmentation model.
• The Faster R-CNN model showed improved polyp detection performance and a reduced false-negative rate with synthetic data augmentation.
• Results on the ETIS-PolypLaribDB dataset surpassed those of similar studies in polyp detection.

Conclusions:

• Controlling the quantity of synthetic and traditional augmentation allows for tuning the sensitivity of deep learning models for polyp analysis.
• GAN-based data augmentation is a practical and effective strategy for enhancing deep learning models in polyp segmentation and detection.