Related Experiment Video
Updated: Aug 16, 2025
06:48
Automated Segmentation of Cortical Grey Matter from T1-Weighted MRI Images
Published on: January 7, 2019
9.0K
Fully Automated Segmentation Models of Supratentorial Meningiomas Assisted by Inclusion of Normal Brain Images
Kihwan Hwang1, Juntae Park2, Young-Jae Kwon3
1Department of Neurosurgery, Seoul National University Bundang Hospital, Seoul National University College of Medicine, Seongnam-si 13620, Gyeonggi-do, Republic of Korea.
Journal of Imaging
|December 22, 2022
Summary
This study enhances brain tumor segmentation by pre-training a U-Net model with glioma data and augmenting it with normal brain MRIs. A novel balanced Dice loss function improved meningioma segmentation accuracy to a Dice score of 0.84.
Area of Science:
- Medical Imaging and Artificial Intelligence
- Neuro-oncology and Computational Pathology
Background:
- Training automatic brain tumor segmentation models requires substantial data, which is often limited for specific tumor types like meningiomas.
- Existing datasets may not adequately represent the diversity needed for robust model generalization.
- Limited clinical data hinders the development of accurate automated segmentation tools for meningioma detection.
Purpose of the Study:
- To develop an effective strategy for training a brain tumor segmentation model with limited meningioma MRI data.
- To improve the segmentation performance by leveraging larger public datasets and incorporating normal brain images.
- To evaluate the impact of pre-training, data augmentation with normal brains, and a novel loss function on segmentation accuracy.
Main Methods:
- A 3D U-Net architecture was employed as the base model for magnetic resonance image (MRI) segmentation.
- The model was pre-trained on the larger public BraTS 2019 glioma dataset and subsequently fine-tuned using a curated dataset of meningioma and normal brain MRIs.
- A novel balanced Dice loss (BDL) function was implemented to enhance the utility of normal brain MRIs during fine-tuning, replacing the conventional soft Dice loss.
Main Results:
- Model pre-training on glioma data improved segmentation performance, increasing the Dice score from a baseline to 0.76.
- Fine-tuning with the inclusion of normal brain MRIs further boosted performance, achieving a Dice score of 0.79.
- The implementation of the balanced Dice loss (BDL) function resulted in the highest Dice score of 0.84, demonstrating significant improvement in meningioma lesion segmentation.
Conclusions:
- The proposed learning strategy, combining model pre-training and augmentation with normal brain images, effectively addresses data limitations for meningioma segmentation.
- The novel balanced Dice loss function significantly enhances segmentation accuracy, outperforming conventional methods.
- This approach shows considerable potential for improving the clinical utility of automated segmentation tools for meningioma detection using MRI.
