Advancing Segmentation of Intracranial Structures in Brain Imaging

Abstract

There are several challenges and crossroads that a researcher encounters and must navigate while conducting studies in the field of anatomical neuroimaging. These include 1) shortage of large datasets for algorithm testing and the limitations associated with small datasets, 2) finding strategies to make established algorithms more efficient, and 3) development of algorithms robust enough to generalize across different hardware and different acquisition settings.

This thesis seeks to address these challenges. It explores the impact of limited datasets on volumetric analyses from Magnetic Resonance Images, introduces innovative approaches for locating intracranial structures and pathologies from Magnetic Resonance (MR) and Computed Tomography (CT) images, and examines how data selection influences analyses and algorithm performance in diverse scenarios.

One aspect of the present research quantifies and proposes strategies to mitigate the influence of biases, confounders, and random variations that frequently arise in brain volumetric analyses with limited datasets. The findings emphasise the effectiveness of specific metrics in accurately distinguishing between healthy and non-healthy subjects even in the presence of bias, confounders, or random variation.

This thesis also introduces a novel method for segmenting the brain from MR images. This method combines segmentation fusion with a marker-controlled watershed transform and utilizes predictions from established segmentation methods as input. Results demonstrate superior performance compared to the conventional segmentation techniques and other meta-algorithms.

In the field of machine learning, the research recommends effective approaches for creating training data with the aim of segmenting intracranial blood from CT images. A neural network developed for this purpose demonstrates potential in segmenting intracranial blood from head CT images while preserving generalizability.

In conclusion, the challenges posed by limited datasets require special considerations as they impact the development of both machine learning and classical image processing methods for segmenting structures within the head. While classical image processing methods may see reduced usage on their own, they will likely be increasingly integrated with machine learning approaches.