Related Experiment Video
Updated: Jul 9, 2025
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
406
Empowering Vision Transformer by Network Hyper-Parameter Selection for Whole Pelvis Prostate Planning Target Volume
Hyeonjeong Cho1,2, Jae Sung Lee2, Jin Sung Kim1
1Department of Radiation Oncology, Yonsei Cancer Center, Heavy Ion Therapy Research Institute, Yonsei University College of Medicine, Seoul 03722, Republic of Korea.
Cancers
|December 9, 2023
Summary
This study optimized vision transformer U-Net (VT U-Net) for prostate planning target volume (PTV) segmentation. Hyper-parameter tuning improved accuracy, demonstrating its necessity for advanced deep learning models in medical imaging.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiotherapy Planning
Background:
- Convolutional Neural Networks (CNNs) like U-Net are used for organ segmentation but struggle with Planning Target Volume (PTV) segmentation.
- Vision Transformers (VT) offer a newer architecture for medical image analysis.
- nnU-Net provides a framework for optimizing network hyperparameters.
Purpose of the Study:
- To improve auto-segmentation of prostate PTV using a Vision Transformer U-Net (VT U-Net).
- To optimize key hyperparameters (patch size, embedded dimension) for VT U-Net.
- To evaluate the performance of hyperparameter-tuned VT U-Net against other deep learning models.
Main Methods:
- Adopted VT U-Net architecture for prostate PTV segmentation.
- Applied nnU-Net-inspired hyperparameter optimization, focusing on patch size and embedded dimension.
- Conducted 4-fold cross-validation on 140 CT scans.
- Compared VT U-Net v.2 performance against seven other deep learning networks.
Main Results:
- The hyperparameter-tuned VT U-Net v.2 achieved an average Dice Similarity Coefficient (DSC) of 82.5% and a 95% Haussdorff Distance (HD95) of 3.5.
- This performance was superior to seven other recently proposed deep learning networks.
- nnU-Net, despite using convolutional layers, also showed competitive results after hyperparameter optimization.
Conclusions:
- Hyperparameter tuning is crucial for optimizing deep learning models, including novel Vision Transformer architectures.
- The optimized VT U-Net v.2 demonstrates significant potential for accurate prostate PTV auto-segmentation in radiotherapy.
- This study highlights the necessity of careful hyperparameter selection for advancing medical image segmentation technologies.
