Data Augmentation using Spatial Transformation for Brain Tumor Segmentation Improvement

Manual segmentation of MRI images, which is important for ultimately improving patient outcomes, is time-consuming, prone to error, and heavily dependent on the expertise of radiologists. To address the challenges of manual segmentation, deep learning based automatic brain tumor segmentation methods are being developed to enhance diagnostic efficiency and accuracy. However, deep learning algorithms typically require extensive training data, which makes data gathering challenging due to the limited occurrence of medical abnormalities such as brain tumors. In this study, we present a spatial transformation-based data augmentation combination method for brain tumor segmentation. By integrating multiple data augmentation methods, we can create a data augmentation pipeline that achieves results that are superior to those of previous studies. Experiment results proves that the proposed data augmentation method improves model prediction's Dice scores to 87.10 for enhancing tumor, 86.89 for tumor core, and 90.91 for whole tumor. The proposed data augmentation method outperformed previous methods even though it employs a simpler architecture. The proposed method's simplicity and lack of need for architectural modifications mean that it can be combined with more complex architectures to further enhance their performance.