Unravelling the effect of data augmentation transformations in polyp segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning models require large annotated datasets, often unavailable in medical imaging.
• Data augmentation is a common strategy to address data scarcity.
• Optimal data augmentation techniques for specific medical imaging tasks remain unclear.

Purpose of the Study:

• To investigate the impact of various data augmentation transformations on polyp segmentation using deep learning.
• To identify the most effective augmentation strategies for different medical imaging datasets.

Main Methods:

• Evaluated image-based (shifts, rotation, shear, zoom, flips, elastic deformation), pixel-based (brightness, contrast), and application-based (specular lights, blurry frames) transformations.
• Trained deep learning models on CVC-EndoSceneStill and Kvasir-SEG datasets with and without augmentation.
• Performed statistical analysis to compare baseline performance against each augmentation strategy.

Main Results:

• Pixel-based transformations (brightness, contrast) significantly improved polyp segmentation in the CVC-EndoSceneStill dataset.
• Image-based transformations, particularly rotation and shear, showed greater benefits for the Kvasir-SEG dataset.
• Augmentation with synthetic specular lights also enhanced performance across datasets.

Conclusions:

• Pixel-based transformations offer significant potential for polyp segmentation in datasets like CVC-EndoSceneStill.
• Image-based transformations are more suitable for datasets such as Kvasir-SEG.
• Dataset characteristics, including polyp size and image properties, influence the effectiveness of different augmentation techniques.