Multimodal Brain Tumor Segmentation Using Transformer and UNET

Abstract

In medical image analysis, accurate 3D brain tissue segmentation is crucial. This study undertakes a meticulous comparative evaluation of two 3D segmentation models: the established 3DUnet and the innovative 3DUnet Transformer. The assessment is anchored on the segmentation accuracy of three critical brain tissue classes; Whole Tumor (WT) Enhancing Tumor (ET) and Tumor Core (TC), with the Dice similarity coefficient (DSC) serving as the principal metric of performance. 3DUnet model demonstrated a consistent advantage over the 3DUnet Transformer, yielding superior DSC scores for ET (0.771 vs. 0.753), WT (0.891 vs. 0.874), and TC (0.804 vs. 0.779). Complementing the DSC, the Hausdorff distance was employed as an additional metric, further evidencing the 3DUnet's enhanced precision in segmentation with lower distance values across the board. These findings highlight the 3DUnet's robust performance and its potential to significantly refine brain tissue segmentation processes within medical imaging applications.