Machine learning based brain tumour segmentation on limited data using local texture and abnormality.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Brain tumour segmentation is complex due to variations in tumour characteristics and imaging parameters.
• Existing methods often require multiple MRI sequences, which are not always available.
• There is a need for automated segmentation techniques that utilize minimal data.

Purpose of the Study:

• To develop a novel automated brain tumour segmentation method using Random Forests.
• To enable accurate segmentation with limited data, specifically contrast-enhanced T1 and FLAIR MRI sequences.
• To evaluate the method's performance on benchmark datasets like BraTS 2013 and BraTS 2017.

Main Methods:

• A Random Forests model was employed, integrating voxelwise texture and abnormality features from two MRI sequences.
• The model generated 275 feature maps to classify voxels into tumour or normal tissue classes.
• A subsequent voxel clustering algorithm refined the segmentation for final tumour delineation.

Main Results:

• The method achieved median Dice scores of 40.9% for low-grade glioma and 75.0% for high-grade glioma active tumour delineation.
• Segmentation accuracy for the total abnormal region, including edema, reached 68.4% and 80.1% respectively.
• High accuracy was observed for contrast-enhancing tissue and edema, with moderate results for non-enhancing tumour tissue and necrosis.

Conclusions:

• The developed automated algorithm effectively segments contrast-enhancing brain tumour components and edema using only post-contrast T1-weighted and FLAIR MRI.
• The approach is particularly suitable for segmenting high-grade gliomas due to its performance on active tumour delineation.
• The method demonstrates the feasibility of accurate brain tumour segmentation with a reduced set of MRI sequences.