Multivariate Image Analysis

Rating:
Summary: The book sets the baseline for the area, and is readable too
Review: If you believe that image sensor technology will continue to be developed concerning size, spatial resolution, measurement accuracy etc, you will have to ask yourself how the increased data volume should be understood, visualized and analysed. The short story on Multivariate Image Analysis (MIA) is that number crunching algorithms are available for data reduction, but the focus is on visualization and the application problem. You will depend on the human ability to explore and iteratively identify the problem, and the ability of the eye-brain to identify relevant information in the visualizations used. I admit that I am a supporter of this strategy.

The book starts out with introducing imaging, images, image operations etc. The authors are well aware that this introduction can not replace the vast literature in the area, it has to be basic, but it is well done. It is interesting to note that the necessity of knowing the characteristics of the image (depending on sensor technology) and the (imaging) experiment is better described here than in conventional image processing literature. A good analysis depend on these factors, right? The main application area magnetic resonance imaging (MRI) is described, but of course it becomes basic too. The used algorithm is principal component analysis (PCA). It is well described, using theory, examples and graphics. I especially appreciate the chapter on pre-processing techniques, with coupling to image and experiment characteristics.

After these introductions, it is time for MIA. The corner stones of MIA; visualisations, data reduction and iterative model work, are described and used in many examples. To be more specific, local models can be created, different matrices can be used, residuals are analysed, a multitude of visualizations are used, and the examples cover many applications (some are a little strange, for example hard bread (kn„ckebr"d)). The main example is an MRI example. The result is a segmentation, or an understanding of the data which helps in further experiments.

The book includes many examples and also high level language code, so it is possible to understand everything in depth if desirable. My own experience is that MIA is quite multi-disciplinary, it demands several experts (sensor technology, image processing, possibly statistics and especially the application expert!) to be successful. Without no do doubt, the book fills the role of creating a common platform for any such project. This book is to my knowledge the only dedicated one presently. There is an historical overview of work in the area which I appreciate very much. The references are adequate, and there are pointers to relevant journals.

Unfortunately, the book can not escape two crucial questions for MIA: What software should be used? After reading about MIA, I would expect a chapter on software. Just think of all the visualisation needed, and what you would like to have in the future. To write your own software would be a never ending project, I know, I have done it. What does the created images show? This is the curse of PCA (and factor analysis in general). Any application expert would wonder, and it therefore becomes a crucial question. The only way I know (and used) to explain this is to use synthetic data, of which you have control, and it is a good exercise to model the image characteristics. This approach is not used, and I can understand that. It puts even more demands on the used software.

Finn.Pedersen (formerly Uppsala University, Sweden)