Factorisation-Based Image Labelling.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Anatomy

Background:

• Accurate segmentation of brain magnetic resonance images (MRI) into anatomical regions is crucial for neuroimaging research.
• Manual segmentation is labor-intensive and costly, necessitating automated solutions.
• Existing automated methods often struggle with variability in image contrasts.

Purpose of the Study:

• To develop a fully automated and general-purpose algorithm for brain MRI segmentation.
• To introduce the Factorisation-based Image Labelling (FIL) model, a novel patched-based label propagation approach.
• To evaluate the FIL model's performance and its ability to handle domain shift across different MRI contrasts.

Main Methods:

• A patched-based label propagation approach utilizing a generative model with latent variables.
• Training the Factorisation-based Image Labelling (FIL) model on a diverse dataset.
• Comparative analysis against state-of-the-art methods using data from the MICCAI 2012 Grand Challenge.
• Assessment of domain shift robustness by segmenting images with varying MR contrasts.

Main Results:

• The proposed Factorisation-based Image Labelling (FIL) model demonstrates effectiveness in segmenting brain anatomical regions.
• The model exhibits robustness in handling different image contrasts.
• Performance comparable to state-of-the-art methods was achieved, with successful handling of domain shift.

Conclusions:

• The Factorisation-based Image Labelling (FIL) model offers a promising automated solution for brain MRI segmentation.
• The model's generalizability across various contrasts and its robustness to domain shift are significant advantages.
• This approach has the potential to streamline neuroimaging analysis and reduce manual annotation efforts.