Unravelling the effect of data augmentation transformations in polyp segmentation

Data augmentation is a common technique to overcome the lack of large annotated databases, a usual situation when applying deep learning to medical imaging problems. Nevertheless, there is no consensus on which transformations to apply for a particular field. This work aims at identifying the effect of different transformations on polyp segmentation using deep learning.

A set of transformations and ranges have been selected, considering image-based (width and height shift, rotation, shear, zooming, horizontal and vertical flip and elastic deformation), pixel-based (changes in brightness and contrast) and application-based (specular lights and blurry frames) transformations. A model has been trained under the same conditions without data augmentation transformations (baseline) and for each of the transformation and ranges, using CVC-EndoSceneStill and Kvasir-SEG, independently. Statistical analysis is performed to compare the baseline performance against results of each range of each transformation on the same test set for each dataset.

This basic method identifies the most adequate transformations for each dataset. For CVC-EndoSceneStill, changes in brightness and contrast significantly improve the model performance. On the contrary, Kvasir-SEG benefits to a greater extent from the image-based transformations, especially rotation and shear. Augmentation with synthetic specular lights also improves the performance.

Despite being infrequently used, pixel-based transformations show a great potential to improve polyp segmentation in CVC-EndoSceneStill. On the other hand, image-based transformations are more suitable for Kvasir-SEG. Problem-based transformations behave similarly in both datasets. Polyp area, brightness and contrast of the dataset have an influence on these differences.