A Patch-Based Approach for the Segmentation of Pathologies: Application to Glioma Labelling

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate brain tumor segmentation is crucial for diagnosis and treatment planning.
• Existing methods often rely on local search windows or require extensive post-processing.

Purpose of the Study:

• To develop a novel, fully-automatic, and generic approach for brain tumor segmentation.
• To enhance the conventional patch-based framework for improved accuracy and efficiency.

Main Methods:

• Utilized multi-atlas patch-based voting techniques for segmentation.
• Improved training data purity and intensity statistics.
• Incorporated augmented features, multi-scale patches, and isometry invariance.
• Employed a probabilistic model for delineating regions of interest.

Main Results:

• Achieved state-of-the-art results on the Multimodal Brain Tumor Image Segmentation challenge datasets.
• Demonstrated highly competitive running times with minimal computational resources.
• Segmentation smoothness was achieved without post-processing.

Conclusions:

• The proposed method offers a robust and efficient solution for fully-automatic brain tumor segmentation.
• The enhancements to the patch-based framework lead to superior performance and reduced risk of overfitting.
• This approach has significant potential for clinical applications in neuro-oncology.