This is more of a booklet than a book, but it is nonetheless a kind of gem for understanding the process of mastering deformable image registration with graphics processing units (GPUs). The 100-plus pages are divided into six chapters. The first chapter introduces the rest of the book, with a brief description of registration, the methods used for registration, the usefulness of deformable registration, and finally the structure of the subsequent chapters.

The book then gets down to business with no concession to superfluous wording or unnecessary prefaces. The authors use a literary technique called in medias res (Latin, meaning “to get in the midst of things”), without describing characters and motivations. Actually, they include a brief description and motivation at the beginning of each chapter, which is very direct and gets right to the point. This approach allows the authors to highlight the optimized registration algorithms, just pointing out the few necessary items the reader needs to follow the logic of the methods. This format is highly suitable for experts in the matter.

The second chapter exposes the unimodal registration of deformable images using B-splines. Unimodal imaging registration refers to the alignment of two or more images taken by the same capture machine. The authors propose a large optimization that can take advantage of the GPU architecture to speed up the method. The approach is subjected to a performance analysis of quality, including sensitivity to volume size and to control point spacing. GPU-optimized versions of the registration method run 39 times faster than a single-core central processing unit (CPU) implementation.

Chapter 3 deals with multimodal registration using B-splines. In this chapter, multimodal registration means the images have been obtained using different imaging methods and thus will have to be matched. In this case, an optimized GPU algorithm runs 28 times faster than the single CPU version.

The fourth chapter explores methods for analytically obtaining the parameters that provide the best regularization of the vector field of images. Vector fields enable the transformation of non-rigid structures to be mapped between images, and provide a measure of the movements of the local substructures over time.

Chapter 5 describes registration by an optical flow method. This chapter is very short and concise. Just a few results are provided, suggesting that the authors prefer the B-spline registration method rather than the optical flow one.

The final chapter introduces Plastimatch, a software package for performing radiotherapy imaging with deformable registration using the GPU-optimized methods described in this book. This software suite is open source and freely downloadable; it was developed by the authors, among others.

In conclusion, many readers will wish for additional chapters, as many registration methods are not described or analyzed. However, the methods included in this book are explained in great detail. I recommend it to experts in the field of deformable image registration and those with deep knowledge of computing and computer architecture (mainly parallel programming with GPUs).

It is regrettable that the images are in grayscale instead of color, which is of course preferable. The authors do refer readers to the color versions of the graphics in the web version of the book, although I was unable to find the uniform resource locator (URL) listed anywhere.