[Paper Review] What is the best data augmentation approach for brain tumor segmentation using 3D U-Net?

Marco Domenico Cirillo, David Abramian+1 more|arXiv (Cornell University)|Oct 26, 2020

Advanced Neural Network Applications11 references4 citations

TL;DR

This study evaluates data augmentation techniques for 3D U-Net-based brain tumor segmentation using the BraTS 2020 dataset. It finds that brightness augmentation and elastic deformation independently yield the best performance, with combined augmentation not providing additional gains over using a single technique.

ABSTRACT

Training segmentation networks requires large annotated datasets, which in medical imaging can be hard to obtain. Despite this fact, data augmentation has in our opinion not been fully explored for brain tumor segmentation (a possible explanation is that the number of training subjects (369) is rather large in the BraTS 2020 dataset). Here we apply different types of data augmentation (flipping, rotation, scaling, brightness adjustment, elastic deformation) when training a standard 3D U-Net, and demonstrate that augmentation significantly improves performance on the validation set (125 subjects) in many cases. Our conclusion is that brightness augmentation and elastic deformation works best, and that combinations of different augmentation techniques do not provide further improvement compared to only using one augmentation technique.

Motivation & Objective

• To investigate the impact of various data augmentation techniques on 3D U-Net performance for brain tumor segmentation.
• To determine whether combining multiple augmentation strategies improves segmentation accuracy beyond using individual techniques.
• To evaluate the effectiveness of augmentation in a relatively large dataset (369 subjects) where data scarcity is less of a concern.
• To provide practical guidance on optimal augmentation use in 3D medical image segmentation tasks.

Proposed method

• Applied standard 3D U-Net for brain tumor segmentation on the BraTS 2020 dataset.
• Implemented five augmentation techniques: flipping, rotation, scaling, brightness adjustment, and elastic deformation.
• Trained separate models using each augmentation type individually on the training set (369 subjects).
• Evaluated model performance on the validation set (125 subjects) using standard segmentation metrics.
• Compared performance across models trained with single augmentation types and those using combinations.
• Selected the best-performing augmentation strategy based on validation set performance.

Experimental results

Research questions

• RQ1 Which individual data augmentation technique yields the highest segmentation performance for 3D U-Net in brain tumor segmentation?
• RQ2 Does combining multiple augmentation techniques lead to further performance improvements compared to using a single technique?
• RQ3 How effective are standard augmentation methods in a dataset with a relatively large number of annotated subjects (n=369) for brain tumor segmentation?
• RQ4 Do certain augmentation types, such as elastic deformation or brightness adjustment, provide more consistent gains than others?

Key findings

• Brightness augmentation produced the most significant improvement in segmentation performance on the validation set.
• Elastic deformation also yielded strong performance gains, ranking among the top augmentation techniques.
• Combinations of multiple augmentation techniques did not result in better performance than using a single technique.
• Flipping, rotation, and scaling showed more modest improvements compared to brightness adjustment and elastic deformation.
• The study confirms that even in large datasets, data augmentation remains a valuable tool for improving 3D U-Net generalization.
• The best results were achieved when using either brightness adjustment or elastic deformation alone, without combining them.