April E. Khademi,
B.Eng., M.A.Sc., Ph.D. in Electrical Engineering
 
http://www.dsp.utoronto.ca/~akhademi/
Multimedia Lab
Dept. of Electrical and Computer Engineering
University of Toronto
 
Research

Thesis Title:
Medical Image Processing Techniques for the Objective Quantification of pathology in MRI

My research concerns medical image processing of cerebral (brain) MRI. I am researching new methods to quantify white matter lesions in the brain (they are neurodegenerative lesions, which reside in the brain’s white matter). See below for example of the lesions in FLAIR MRI (they have a bright appearance). As white matter lesions are a precursor of future stroke, understanding earlier stages of the disease can lead to better intervention protocols and therapy strategies. My main objective is to arrive at novel measures that describe the distribution of these lesions throughout the white matter. This research is being conducted in collaboration with Sunnybrook Hospital (Medical Imaging Research Department).

Examples of cerebral MRI with periventricular white matter lesions.

Of interest to the medical community is the volume of these lesions. Traditionally, to compute the volume, the radiologist would have to manually outline each lesion carefully (for several patients and images). This is not only labourious, time consuming and error prone, but it is also subjective. To combat these downfalls, automated segmentation can be used for objective, reliable and efficient results. This is the main focus of this research: to develop a robust and accurate white matter lesion segmentation scheme.

The main challenge of segmentation in neuro MRI is that there are at least three types of artifacts that severely impede automated approaches, namely: acquisition noise, inhomogeneity bias field, and partial volume averaging. These degradations change the distribution of the image in a manner that can be difficult to model and thus reduces the performance of automatic algorithms. The partial volume averaging artifact is a very important distortion to be considered when computing the volume of the lesions. It concerns the way a voxel is imaged, when there is more than one tissue type found within the extent of the volume element. The tissue in these regions is known as “mixture” tissue since it is composed of more than one tissue type.

To measure the lesions accurately, quantification of the partial volume averaging effect is critical in determining how much of each tissue is found in these mixture pixels. For this, we have developed a new partial volume quantification scheme that is image-based and is built on mathematical principles surrounding the physics of PVA. It does not require any parameters, distributional assumptions or any other apriori information. This is an advantage over the other approaches, which need initialization parameters, training samples, or distributional assumptions. An sample of the results are shown below. The top row contains the manual (expert) segmentations, and the bottom row contains the lesion classification maps (red indicates pure lesion).

Manual and automatic segmentation of white matter lesions.



© 2017 April E. Khademi