GLISTR: glioma image segmentation and registration

Area of Science:

• Medical Imaging
• Computational Anatomy
• Neurosurgery

Background:

• Accurate segmentation and registration of brain tumors in MRI are crucial for diagnosis and treatment planning.
• Existing methods often struggle with precise delineation of glioma and surrounding edema.
• Probabilistic atlases aid in understanding population variability but require adaptation to individual scans.

Purpose of the Study:

• To develop a generative approach for simultaneous glioma segmentation and probabilistic atlas registration in brain MR scans.
• To adapt a healthy population atlas to individual patient scans, incorporating tumor and edema characteristics.
• To validate the method against expert segmentations and state-of-the-art techniques.

Main Methods:

• Expectation-maximization (EM) algorithm with a glioma growth model for atlas seeding and adaptation.
• Iterative refinement of tissue label probabilities, deformation fields, and tumor growth parameters.
• Generative approach for simultaneous segmentation, atlas registration, and low-dimensional patient scan description.

Main Results:

• The proposed method achieved superior segmentation accuracy for tumor and edema compared to reference methods.
• Segmentation performance was comparable to a second human rater.
• The method successfully registered probabilistic atlases and provided a low-dimensional tumor description.

Conclusions:

• The generative approach offers an effective solution for automated brain tumor segmentation and atlas registration.
• The method provides accurate delineation of glioma and edema, aiding clinical interpretation.
• The study demonstrates the utility of adaptive probabilistic atlases in neuroimaging research and clinical applications.