Intelligent methods for automatic classification of medical images / Mohammad Reza Zare

Zare, Mohammad Reza (2013) Intelligent methods for automatic classification of medical images / Mohammad Reza Zare. PhD thesis, University of Malaya.

Preview

PDF (Full chapters) - Submitted Version Download (3393Kb) | Preview

Abstract

The ever increasing number of medical images in hospitals urges on the need for generic image classification systems. These systems are in an area of great importance to the healthcare providers. However, classification of large medical database is not an easy task due to unbalance number of training data, intra class variability and inter-class similarities among them. In this thesis, three lassification frameworks are presented to increase the accuracy rate of every individual categories of such database. Bag of Words (BoW) has been explored for image representation techniques in the proposed frameworks. In the first framework, we proposed an iterative filtering scheme on the database where classes with optimal accuracy rate are filtered out. They are then used to construct a new classification model. These processes are carried out in four iterations. As a result, four classification models are generated from different number of classes. These models are then employed to classify unseen test images. In continuation of the first framework, another classification framework is proposed for classes which are left with low accuracy rate after the first iteration. These classes are those with high ratio of intra class variability and inter-class similarities. The classification process is carried out by employing three different annotation techniques, i.e. Annotation by binary classification, Annotation by Probabilistic Latent Semantic Analysis (PLSA) and Annotation using top similar images. The annotated keywords produced are integrated by applying ranking similarity. The final annotation keywords were then divided into three levels according to the body region, specific bone structure in body region as well as imaging direction. Different weights were given to each level of the keywords; they are then used to calculate the weightage for each category of medical images based on their ground truth annotation. The weightage computed from the generated annotation of test iii image was compared with the weightage of each category of medical images, and then the test image would be assigned to the category with closest weightage to the test image. In the third framework, the unsupervised latent space model is used in feature extraction to discover patterns of visual co-occurrence. In this direction, we employed PLSA to learn the co-occurrence information between elements in the vector space. PLSA model can generate a robust, high level representation and low-dimensional image representation. This would help to disambiguate visual words. Thus, a classification framework based on integration of PLSA and discriminative Support Vector Machine (SVM) classifier is developed. In this framework, both visual features and textual features of the images are incorporated by using multi-modal PLSA model. The experimental results on all the above frameworks have shown an increment in accuracy rate at the entire database level as well as at class specific level compared with other methods.

Actions (For repository staff only : Login required)