Classification and Clustering

• Published in: Guide to Medical Image Analysis pp. 473 - 528
• Main Author: Toennies, Klaus D
• Format: Book Chapter
• Language: English
• Published: United Kingdom Springer London, Limited 01.01.2017; Springer London
• Series: Advances in Computer Vision and Pattern Recognition
• Subjects: Decision Boundary; Feature Space; Independent Component Analysis; Support Vector Machine; Visual Word
• ISBN: 9781447173182; 144717318X
• ISSN: 2191-6586; 2191-6594
• DOI: 10.1007/978-1-4471-7320-5_12

Assigning semantics to segments is required if segmentation has not been combined with object detection. Classification is then based on evaluating segment attributes such as shape and appearance. The dimension of feature space is often high (>10) and the number of samples to train a classifier or to deduce a clustering is low. Methods are different compared to classification or clustering of pixels or voxels. For the most part, likelihood functions are not estimated and the classification criterion is directly based on the training data. Feature reduction techniques, classifiers, and clustering methods that focus on analysis in sparse feature spaces are the topic of this chapter. Among the different methods treated in this chapter, deep convolutional neural networks stand out, as their variable applicability has been found widespread use that encompasses applications on a pixel basis. The methods presented here complement the methodology presented in Chap. 10.1007/978-1-4471-7320-5_7.