Effective data augmentation for brain tumor segmentation

This research is to propose a training strategy for 2D U‐Net is proposed that uses selective data augmentation technique to overcome the class imbalance issue. This also helps in generating synthetic data for training which improves the generalization capabilities of the segmentation network. The training data are prepared with random sampling to further reduce the class imbalance. The post‐processing stage is used to decrease the outliers in the final output. The performance of the proposed solution is tested on the online leaderboard. The results achieved on the validation set of Brain Tumor Segmentation 2019 dataset were 0.79, 0.89, and 0.8 for enhancing tumor (ET), whole tumor (WT), and core tumor (CT) respectively. The part of the training set is also evaluated locally, and the results show the effectiveness of using selective data augmentation and random sampling. The multi‐view fusion improved the robustness and overall dice scores.