Image databases in medical applications

Abstract

The number of medical images acquired yearly in hospitals increases all the time. These imaging data contain lots of information on the characteristics of anatomical structures and on their variations. This information can be utilized in numerous medical applications. In deformable model-based segmentation and registration methods, the information in the image databases can be used to give a priori information on the shape of the object studied and the gray-level values in the image, and on their variations. On the other hand, by studying the variations of the object of interest in different populations, the effects of, for example, aging, gender, and diseases on anatomical structures can be detected.

In the work described in this Thesis, methods that utilize image databases in medical applications were studied. Methods were developed and compared for deformable model-based segmentation and registration. Model selection procedure, mean models, and combination of classifiers were studied for the construction of a good a priori model. Statistical and probabilistic shape models were generated to constrain the deformations in segmentation and registration so that only the shapes typical to the object studied were accepted. In the shape analysis of the striatum, both volume and local shape changes were studied. The effects of aging and gender, and also the asymmetries were examined.

The results proved that the segmentation and registration accuracy of deformable model-based methods can be improved by utilizing the information in image databases. The databases used were relatively small. Therefore, the statistical and probabilistic methods were not able to model all the population-specific variation. On the other hand, the simpler methods, the model selection procedure, mean models, and combination of classifiers, gave good results also with the small image databases. Two main applications were the reconstruction of 3-D geometry from incomplete data and the segmentation of heart ventricles and atria from short- and long-axis magnetic resonance images. In both applications, the methods studied provided promising results. The shape analysis of the striatum showed that the volume of the striatum decreases in aging. Also, the shape of the striatum changes locally. Asymmetries in the shape were found, too, but any gender-related local shape differences were not found.