Author: Sang Hee Ahn, Nalee Kim, Do Hoon Lim 👨🔬
Affiliation: Samsung Medical Center, Department of Radiation Oncology, Samsung Medical Center, Sungkyunkwan University School of Medicine 🌍
Purpose: MRI offers superior soft-tissue contrast, aiding tumor localization and segmentation in radiation therapy, which traditionally relies on oncologists' expertise. This study compares CNN-based models and Vision Mamba-UNet for brain tumor segmentation, focusing on their clinical applicability.
Methods: We evaluated the auto-segmentation performance on T1-weighted MRI sequences from 94 patients diagnosed with high grade glioma treated with radiation therapy between 2019 and 2021 at Samsung Medical Center. The dataset was divided into 74 cases for training and 20 cases for testing. Each model was trained under identical conditions and evaluated on the test set. The neural network models assessed included CNN-based models such as UNet, UNet+SimAM (with attention module), ResUNet++, DUCK-Net, UNet3Plus, nnUNet, SAM2-Unet, and the Vision Mamba-based model H-vmunet. Segmentation performance was quantitatively assessed using metrics including the Dice Similarity Coefficient (DSC), 95th percentile Hausdorff Distance (HD95%), and Relative Volume Difference (RVD). Qualitative evaluation was conducted by a radiation oncologist who reviewed the predicted contours and rated them using a Likert scale. The study protocol was approved by the Samsung Medical Center IRB (No. 2024-12-139-001) with a retrospective data collection design.
Results: The quantitative performance metrics for each model on the test set were as follows: DSC values of 0.74, 0.79, 0.76, 0.77, 0.80, 0.87, and 0.81; HD95% values (mm) of 12.29, 16.71, 15.91, 7.87, 17.25, 6.39, and 1.11; and RVD values of 0.34, 0.17, 0.27, 0.18, 0.17, 0.08, and 0.21, corresponding to UNet, UNet+SimAM, ResUNet++, DUCK-Net, UNet3Plus, nnUNet, and H-vmunet, respectively.
Conclusion: This study highlights the potential of advanced deep learning models like H-vMUNet in improving MRI-based brain tumor segmentation accuracy and consistency. While H-vMUNet performed slightly below nnUNet, it surpassed other CNN models and demonstrated strong clinical applicability. These findings suggest its potential to enhance precision and efficiency in adaptive radiation therapy, benefiting patient outcomes.