Abstract
The constant development of medical imaging technology and the dependence of medical professionals
on medical images as an auxiliary data continues to increase the demand for more advanced and
practical solutions that can be used to analyse the information in these images. Due to the wide variety
of information present, its subtle nature, and the need for highest accuracy, analysis of medical images
provide an excellent test case to many image analysis algorithms. Computer aided diagnostic (CAD)
systems are developed considering these factors. Often, CAD systems comprise of multiple algorithms
concatenated in a pipeline to perform different operations. The final output is aimed to be interpretable
by a medical professional.
In this thesis, three image processing techniques - i) enhancement, ii) segmentation and iii) shape analysis
are studied. A CAD system is built using these three operations forming different blocks of pipeline.
The intent of this system is to aid in diagnosis of a specific variant of dementia known as Alzheimer’s
disease (AD) by detecting the stage of an object of interest, a sub-cortical structure called hippocampus.
AD is a neurodegenerative disease with crippling effect on human mind and yet, there is no known
cure to this disease. Hence, the detection of its early onset proves to be helpful to monitor and regulate
its progress. Quantitative analysis performed on segmented volume of hippocampus has been used to
classify it into one of the three stages of diagnosis - normal, mildly affected by the disease and severely
affected by the disease. In order to study the structural changes in sub-hippocampal regions with respect
to progression of the disease, the piecewise quantitative information has been used. Segmentation of
hippocampus is performed by a physics-based model that operates on local and global image forces
to evolve a surface. A coarse segmentation method is used to obtain an estimate of the surface to be
used in evolution. Since the segmentation method used here is dependent on intensity-based features,
image contrast enhancement is performed by an adaptive intensity windowing algorithm which forms
the preprocessing stage to segmentation. Thus, each block operates on the output of the previous block,
and the output of the final block is useful information to diagnosis.
The evaluation of intensity windowing is done by comparing the quantitative measures of accuracy of
segmentation algorithm on non-windowed and windowed images from one of the publicly available
datasets. The evaluation of segmentation is done by conducting experiments on images from two publicly
available datasets and by comparing its accuracy with those from other recent methods on the
same datasets. Also, to demonstrate its general applicability, three other structures have been segmented
using this segmentation method. Images and their corresponding diagnostic information, belonging to
three different stages, have been used from another publicly available dataset for quantitative analysis
of hippocampus. The segmentation is performed by the proposed method and its output is used for
classification into different stages. Encouraging results have been observed at each of these stages. That
is, the proposed method does not perform exceedingly better than some methods, but does so than some
others. With further investigation to address the pitfall of the proposed method, one could expect to
achieve results to match the state of the art, retaining the advantages of this method.
The proposed segmentation method has been tested on 43 volumes drawn from two publicly available
datasets and by measuring Dice coefficient with reference to hand segmented ground truth, hippocampus
was segmented with an average accuracy of 77%, amygdala with an average accuracy of 68%, caudate
with an average accuracy of 82% and putamen with an average accuracy of 80%. The classification
scheme has been assessed on 150 volumes from the ADNI database and the normal hippocampus was
differentiated from the abnormal cases with an overall accuracy of roughly 80%.