Related Experiment Video
Updated: Jul 9, 2025
08:41
Patient-Specific Polyvinyl Alcohol Phantom Fabrication with Ultrasound and X-Ray Contrast for Brain Tumor Surgery Planning
Published on: July 14, 2020
8.5K
RFS+: A Clinically Adaptable and Computationally Efficient Strategy for Enhanced Brain Tumor Segmentation.
Abdulkerim Duman1, Oktay Karakuş2, Xianfang Sun2
1School of Engineering, Cardiff University, Cardiff CF24 3AA, UK.
Cancers
|December 9, 2023
Summary
This study introduces a Region-Focused Selection Plus (RFS+) strategy to improve deep learning models for brain tumor segmentation. RFS+ enhances model generalization and quantification, requiring less data and time for training.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computational Biology
Background:
- Automated brain tumor segmentation is crucial for diagnosis and treatment planning using MRI modalities (T1, T1ce, T2, FLAIR).
- Current deep learning models excel on standardized datasets but face challenges in diverse clinical settings due to variations in image acquisition parameters.
- A need exists for robust segmentation models that generalize well across different clinical environments.
Purpose of the Study:
- To introduce and evaluate the novel Region-Focused Selection Plus (RFS+) strategy for improving deep learning-based automatic brain tumor segmentation.
- To enhance the generalization and quantification capabilities of deep learning models in diverse clinical settings.
- To reduce the computational resources and training time required for accurate brain tumor segmentation.
Main Methods:
- Developed the Region-Focused Selection Plus (RFS+) strategy, a targeted approach focusing on individual regions with customized input masks, activation functions, loss functions, and normalization.
- Employed weighted ensemble learning by identifying top-performing models for specific regions.
- Investigated multi-class, multi-label, and binary segmentation approaches with various normalization techniques, comparing three U-net variants on the BraTS 2021 validation dataset and a local dataset.
Main Results:
- The 2D U-net model achieved Dice Similarity Coefficient (DSC) scores of 77.45% (ET), 82.14% (TC), and 90.82% (WT) on the BraTS 2021 validation dataset.
- The 2D U-net model augmented with RFS+ strategy achieved a superior DSC score of 79.22% for gross tumor volume (GTV) on the local dataset.
- The RFS+ strategy-enabled model required 10% less training data, 67% less memory, and 92% less training time compared to state-of-the-art models.
Conclusions:
- The RFS+ strategy effectively enhances the generalizability and quantification of deep learning models for brain tumor segmentation.
- RFS+ offers a computationally efficient approach, reducing data, memory, and time requirements.
- This strategy holds significant promise for improving clinical applications of automated brain tumor segmentation.
