Automated brain tumour detection and segmentation using superpixel-based extremely randomized trees in FLAIR MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neuro-oncology

Background:

• Brain tumors require accurate detection and segmentation for effective patient management.
• Manual delineation of brain tumors from MRI is time-consuming and subjective.
• Automated methods can improve efficiency and reproducibility in brain tumor analysis.

Purpose of the Study:

• To develop a fully automated method for detecting and segmenting brain tumors (tumor core and edema) in FLAIR MRI.
• To compare the performance of Extremely Randomized Trees (ERT) and Support Vector Machine (SVM) classifiers for superpixel classification.

Main Methods:

• A superpixel-based classification approach is employed.
• Novel image features including intensity, Gabor textons, fractal analysis, and curvatures are extracted from superpixels.
• Extremely Randomized Trees (ERT) and Support Vector Machine (SVM) classifiers are used to differentiate tumor from non-tumor regions.

Main Results:

• The automated method demonstrated high detection and segmentation performance on clinical and BRATS datasets.
• For the clinical dataset, average detection sensitivity was 89.48%, balanced error rate 6%, and Dice overlap 0.91.
• For the BRATS dataset, results showed 88.09% sensitivity, 6% error rate, and 0.88 Dice overlap.

Conclusions:

• The proposed automated method closely matches expert delineation across all glioma grades.
• This technique offers a faster and more reproducible approach to brain tumor detection and delineation.
• The method has the potential to significantly aid in patient management for brain tumor cases.