Abstract
In order to diagnose a cancer, medical imaging techniques like radiography, ultrasound
or magnetic resonance are commonly used. By means of these techniques
is possible to detect regions with high risk of containing cancer cells, whose diagnostic
must be conrmed by getting a cell sample (biopsy). However, these kind of
images are not easy to interpret, which causes that, sometimes, the physician is not
able to detect the presence of a cancer (false negative), even when he or her has a
large experience.
One possible way of improving the diagnosis and reducing the number of false
negatives is to use computer-aided diagnosis (CAD). A CAD system analyses the
medical image and tries to detect suspicious regions. These regions are highlighted
in order to draw the attention of a human expert and to provide a second read of
the image.
In this thesis, different methods of computer vision and pattern recognition aimed
at cancer detection from medical images are presented and tested. The goal is
to implement a CAD system that can help to improve the diagnosis. The study is
focused on prostate cancer diagnosis from ultrasound images and on breast cancer
diagnosis from mammographies. Different extraction features methods based on
intensity values, textures and gradient have been tested. A non-parametric classier
based on distances (k-nearest neighbors) and a parametric classier based on Markov
Models have been used. Also, Support Vector Machines (SVM) and a Boosting
classier have been tested. The different difculties that arise in computer-aided
diagnosis are explained and solutions for each one are proposed.
Prostate computer-aided diagnosis is an extremely difcult task and with very
few studies have been published in this area. In spite of the difculties, promising
results indicating that valid information for the diagnosis is present in the image
have been obtained. On the other hand, breast computer-aided diagnosis has been
researched in depth and many studies related to this topic have been published. The
results obtained in this thesis have been encouraging, and are similar, or better in
some cases, than other previously published.