Related Experiment Video
Updated: Dec 28, 2025
10:25
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
49.2K
Improving Patch-Based Convolutional Neural Networks for MRI Brain Tumor Segmentation by Leveraging Location
Po-Yu Kao1, Shailja Shailja1, Jiaxiang Jiang1
1Vision Research Lab, Department of Electrical and Computer Engineering, University of California, Santa Barbara, Santa Barbara, CA, United States.
Frontiers in Neuroscience
|February 11, 2020
Summary
This study introduces a novel method for automated brain tumor segmentation by integrating location information with neural networks. This approach enhances segmentation accuracy for improved diagnostic tools.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuroscience
Background:
- Manual brain tumor annotation is time-consuming and resource-intensive.
- Accurate automated segmentation tools are crucial for clinical applications.
- Brain tumor lesions exhibit non-uniform distribution across brain regions.
Purpose of the Study:
- To develop a novel method for automated brain tumor segmentation.
- To integrate location information with patch-based neural networks for improved accuracy.
- To enhance segmentation performance by leveraging brain parcellation data.
Main Methods:
- Utilized a brain parcellation atlas in the Montreal Neurological Institute (MNI) space, mapped to individual subject data.
- Integrated mapped atlas information with structural Magnetic Resonance (MR) imaging data.
- Trained patch-based neural networks (3D U-Net, DeepMedic) and employed a two-level ensemble method with XGBoost fusion.
Main Results:
- The proposed location information fusion method significantly improved segmentation performance of 3D U-Net and DeepMedic.
- The ensemble method achieved superior segmentation performance compared to state-of-the-art networks on the BraTS 2017 dataset.
- The approach rivaled state-of-the-art performance on the BraTS 2018 benchmark dataset.
Conclusions:
- Integrating location information with neural networks offers a promising approach for accurate automated brain tumor segmentation.
- The developed two-level ensemble method effectively combines the strengths of different neural network models and location data.
- This method has the potential to improve clinical workflows and patient outcomes in neuro-oncology.
