You are here: Machine Learning for Precision Health»Parcellation of Diseased Brains

Parcellation of Diseased Brains

Segmentation of Diseased Brains Using Limited Training Data

Accurate labeling of anatomical structures (parcellation) is a fundamental diagnostic tool to quantify diseases and disorders with characteristic morphological changes such as schizophrenia, Alzheimer’s, Huntington’s, Parkinson’s, hydrocephalus, and developmental disorders.

It is also a key enabling tool for quantitative assessment of brain connectivity and regional metabolic activity.

However, standard atlas-based parcellation approaches do not provide accurate labeling when labeled training data is in limited supply, a situation which occurs frequently (e.g. when an imaging method is updated or a new patient cohort is studied).

Parrellation Diagram

Comparison of parcellation results on the T1 weighted images of brains of Alzheimer’s patients. The first column shows an axial slice from the volume of two subjects with advanced Alzheimer’s, while the second column shows the manual labeling from expert neuroanatomists. Our result (right column) accurate labels structures throughout the brain and substantially improved over other methods such as FreeSurfer and iSTAPLE.
 

We develop methods that provide the accurate parcellations for such cases by fusing the multi-atlas and probabilistic atlas approaches. In application to Alzheimer’s brains, my approach labels 34 structures with significantly more accuracy than other widely used methods (Figure) and more accurately than the best methods from the MICCAI 2012 challenge.

Our current work aims to extend the approach to brains with tumors and contusions.