Classification of medical images is a fundamental step in different applications as e.g. within a medical image retrieval system . Due to the high variability of medical image data it is important to use appropriate models in the classification process . We describe a classification method for medical images based on non-linear image distortion models that considerably improves classification results with respect to other known methods on a database of medical radiographs. The most important improvement that leads to a significant reduction of the error rate is the inclusion of feature vectors using image gradient and image parts as local context instead of the value of one pixel only.
A large variety of methods for classification of medical images is discussed in the literature. A number of these have also been evaluated on the data used in this work, the RWTH Aachen University IRMA (image retrieval in medical applications) database. Best results on these data were achieved using a statistical model incorporating various techniques that cope with the inherent variability of the data . Other techniques like the use of cooccurrence matrices or the Euclidean nearest neighbor yielded higher error rates when applied to this task. The statistical approach with a model of variability (distorted tangent distance) obtained an error rate of 8.0% on the used database of radiographs.
We propose the use of a distortion model for classification, which is connected to the field of image registration by the inherent optimization or matching process. The topic of image registration is a wide research area especially in the domain of medical image processing and elaborate techniques exist. The fundamental difference to the matching methods that result from an image distortion model is the objective of the matching process: In distortion modeling for classification the aim is to compensate only those deformations that leave the class unchanged. Deformations that change the class are unwanted in the matching, i.e. the emphasis is placed on discrimination between classes. On the other hand in image registration it is usually known that the images are from the same class (e.g. the same body region of the same patient) and the best matching is sought in order to determine the differences between the two images.