Feature Extraction, Shape Fitting and Image Segmentation

This lecture covers the related topics of feature extraction, shape fitting and image segmentation. Just about all quantitative analysis of medical images requires some form of segmentation or feature extraction. Segmentation [3] [12] [13] distinguishes structures, regions or tissue classes of interest from other detail in the images. Feature extraction can be used to identify specific structures (e.g. points, blobs, curves, edges, surfaces etc) which often have biological importance e.g. organ or lesion boundaries, vessels etc. Some techniques begin with feature extraction to identify points of interest then build a generic shape model using those features as a reference and finally use the fitted shape model to perform image segmentation on new (i.e. previously unseen) images. Once structures are labelled, quantitative information about volume or shape can be extracted and comparisons can be made. The most common useful comparisons are in the same subject over time (e.g. to track growth) or between groups of subjects (e.g. to identify systematic structural brain differences between two groups of subjects who might score differently in psychological testing). I will first briefly review some segmentation techniques that can be applied to distinguish tissue classes and then focus on methods of fitting shapes to medical images that include some knowledge of either the image properties or the population variation of the structure, or both. These methods are generically known as deformable models. A somewhat artificial division can be drawn between geometric deformable models and statistical shape models. The former are geometric models that evolve to match an instance of a structure. The evolution is driven by forces that are typically functions of the local image environment and constrained by prior knowledge often in the form of simple geometric constraints. Statistical shape models inherently require training data and produce compact representations of structural variation across a population. A wide range of plausible structure can be generated by varying a small number of parameters, resulting in an efficient optimisation task when fitting to new data. There has been a vast amount of research in all the methods mentioned in these notes and to do justice to these topics would require a book. Therefore please use this material as pointers to further reading rather than a definitive account. Errors and omissions are my responsibility.