- English
- français

# Distance transformations: fast algorithms and applications to medical image processing

Medical image processing is a demanding domain, both in terms of CPU and memory requirements. The volume of data to be processed is often large (a typical MRI dataset requires 10 MBytes) and many processing tools are only useful to the physician if they are available as real-time applications, i.e. if they run in a few seconds at most. Of course, a large part of these demands are - and will be - handled by the development of more powerful hardware. On the other hand, when faced with non-linear computational complexity, the development of improved algorithms is obviously the best solution. Distance transformations, a powerful image analysis tool used in a number of problems such as image registration, requires such improvements. A distance map is an image where the value of each pixel is the distance from this pixel to the nearest pixel belonging to a given set or object. A distance transformation (DT) is an algorithm that computes a distance map from a binary image representing this set of pixels. This definition is global in the sense that it requires finding the minimum on a set of distances computed between all image pixels and all object pixels. Therefore, a direct application of the definition usually leads to an unacceptable computational complexity. Numerous algorithms have been proposed to localize this definition of distance to the nearest pixel and allow a faster DT computation, but up to now, none of them combines both exactness and linear complexity. Numerous applications of distance transformations to image analysis and pattern recognition have been reported and those related to medical image processing are explored in what follows. Chapter 1 introduces a few basic concepts, a typical application of distance transformations in pattern recognition and the key challenges in producing a DT algorithm. Chapter 2 contains an exhaustive critical review of published algorithms. The strong and weak points of the most popular ones are discussed and the core principles for our original algorithms are derived. Chapters 3, 5, 6, 8 and 10 present original distance transformation algorithms. Each of those chapters is organized in a somewhat similar fashion. First we describe the algorithm. Then we evaluate its computational complexity and compare it to the state of the art. Chapter 4, 7, 9 and 11 each present an application to a particular problem in medical image processing, using the algorithm developed in the previous chapter. Ideally, the description of any medical image processing problem should include a medical justification of the need for an automated processing, a complete review of the state of the art in the field, a detailed description of the proposed processing method, and an evaluation of the accuracy of the results and their medical significance. Because of both time and space constraints in this thesis, such an exhaustive work will only be presented for the application in chapter 4, while the other applications will be described more briefly. Chapter 3 describes a new exact Euclidean distance transformation using ordered propagation. It is based on a variation of Ragnelmam's approximate Euclidean DT. We analyze the error patterns for approximate Euclidean DT using finite masks, and we derive a rule defining, for any pixel location, the size of the neighborhood that guarantees the exactness of the DT. This algorithm is particularly well-suited to implement mathematical morphology operations, which are examined in details. In Chapter 4, we apply the algorithm of chapter 3 to the segmentation of neuronal fibers from microscopic images of the sciatic nerve. In particular, it is used to determine the thickness of the myelin sheath surrounding the center of the fiber. This study was carried out in collaboration with the Neural Rehabilitation Engineering Laboratory, UCL. Chapter 5 proposes another exact Euclidean distance transformation, based on the explicit computation of the Voronoi division of the image. Possible error locations are detected at the corners of the Voronoi polygons and corrected if needed. This algorithm is shown to be the fastest exact EDT to date. It approaches the theoretical optimal complexity, a CPU time proportional to the number of pixels on which the distance is computed. Chapter 6 investigates how the algorithms of chapters 3 and 5 can be extended to 3 dimensional images. It shows the limitations of both approaches and proposes an hybrid algorithm mixing the method of chapter 5 and Saito's. In Chapter 7, the 3D Euclidean DT is applied to the registration of MR images of the brain where the matching criterion is the distance between the surfaces of similar objects (skin, cortex, ventricular system, ...) in both images. Examples are shown, from projects with the Neuro-physiology Laboratory, UCL, and with the Positron Tomography Laboratory, UCL. Chapter 8 discusses an extension of the distance transformation concept: geodesic distances on non-convex domains. Because geodesic distances are based on the notion of paths, a trade-off has to be introduced between the accuracy with which straight lines are represented and the way curves of the domain are followed. It is shown that, whatever the trade-off chosen, there is an efficient implementation of the geodesic DT by propagation. By back-tracking the geodesic distance propagation, one can find the shortest path between a target and a starting point. In chapter 9, this is used to plan the optimal path for the camera movements in virtual endoscopy, a work done in collaboration with the Surgical Planning Laboratory, Harvard Medical School, Boston. Chapter 10 extends the Euclidean distance transformation from finding the nearest object pixel to finding the k nearest object pixels. It is shown that this can be done with a complexity increasing linearly with k. In Chapter 11, the k-DT is used as a fast implementation of the k Nearest Neighbors (k-NN) classification between different tissue types in multi-modal MR imaging. This is illustrated through the classification of multiple sclerosis lesions from T1-T2 images, provided by the Radiology unit, St-Luc Hospital, UCL, via the Positron Tomography Laboratory, UCL. Finally, a general conclusion is drawn. It reviews the main contributions of the thesis, its applications and explores some new domains in which their applications could also be useful. Ultimately, the publications related to this thesis are briefly reviewed.

Mots-clefs: Euclidean Distance Transform

#### Référence

Notice créée le 2005-11-18, modifiée le 2016-08-08