Comparative Analysis of Brain Tumor Image Segmentation Performance of 2D U-Net and 3D U-Nets with Alternative Normalization Methods

Abstract

Advancements in deep learning-based brain tumor image segmentation have significantly contributed to the rapid and accurate diagnosis of brain tumors. U-Net, a deep learning model used for brain tumor image segmentation, serves as the basic architecture for many such models. Although U-Nets are categorized into two-dimensional (2D) and three-dimensional (3D) models, it remains unclear which model is more effective for brain tumor image segmentation. Therefore, this study compared the performances of 2D and 3D U-Net models for brain tumor image segmentation. In this study, for 2D U-Net, we employed batch normalization. For the 3D models, three variants with distinct normalization techniques were used: 3D BN U-Net with batch normalization, 3D GN U-Net with group normalization, and 3D IN U-Net with instance normalization. The dataset consisted of brain MRI images from 600 patients with brain tumors and expert-labeled mask images. Experiments were conducted using 5-fold cross-validation, and the results revealed that the 3D GN and IN models outperformed the 2D model. In conclusion, for U-Net models in brain tumor image segmentation, the 3D GN and IN U-Net models, which replaced batch normal ization with group normalization or instance normalization, proved to be the most effective