This thesis focuses on voxel/pixel classification based approaches for image
segmentation. The main application is segmentation of articular cartilage
in knee MRIs. The first major contribution of the thesis deals with large
scale machine learning problems. Many medical imaging problems need huge
amount of training data to cover sufficient biological variability. Learning
methods scaling badly with number of training data points cannot be used in
such scenarios. This may restrict the usage of many powerful classifiers having
excellent generalization ability. We propose a cascaded classifier which
allows usage of such classifiers in large scale problems. We demonstrate its
application for segmenting tibial articular cartilage in knee MRI scans, with
number of training voxels being more than 2 million. In the next phase
of the study we apply the cascaded classifier to a similar but even more
challenging problem of segmenting femoral cartilage. We discuss similarities
and provide our solutions to the challenges. Our cascaded classifier for
cartilage segmentation comprised of two stages of classification combining
nearest neighbour classifier and support vector machine. We compared our
method to a state-of-the-art method for cartilage segmentation using one
stage nearest neighbour classifier. Our method achieved better results than
the state-of-the-art method for tibial as well as femoral cartilage segmentation.

The next main contribution of the thesis deals with learning features
autonomously from data rather than having a predefined feature set. We
explore deep learning approach of convolutional neural network (CNN) for
segmenting three dimensional medical images. We propose a novel system
integrating three 2D CNNs, which have a one-to-one association with the
xy, yz and zx planes of 3D image, respectively and this system is referred
as triplanar convolutional neural network in the thesis. We applied the
triplanar CNN for segmenting articular cartilage in knee MRI and compared
its performance with the same state-of-the-art method which was used as
a benchmark for cascaded classifier. Although our method used only 2D
features at a single scale, it performs better than the state-of-the-art method
using 3D multi-scale features. In the latter approach, the features and the
classifier have been carefully adapted to the problem at hand. That we were
able to get better results by a deep learning architecture that autonomously
learns the features from the images is the main insight of this study.

While training the convolutional neural networks for segmentation purposes,
the commonly used cost function does not consider the labels of the
neighbourhood pixels/voxels. We propose spatially contextualized convolutional
neural network (SCCNN) which incorporates the labels of the neighbouring
pixels/voxels while training the network. We demonstrate its application
for the 2D problem of segmenting horses from the Weizmann horses
database using 2D CNN and our 3D problem of segmenting tibial cartilage
in knee MRIs using triplanar CNN. The proposed SCCNN improved the
segmentation performance in both the cases. The good results obtained by
SCCNN encourage to gain more insight into such frameworks.