Malign lymphoma are the seventh common death cause in the western hemisphere. The therapy of the patients and the prognosis are highly related to the dispersal pattern of the lymphom node disease. For this reason imaging diagnostics of the whole body are required on a regular basis. Prospectively whole-body magnetic resonance imaging (MRI) will gain more and more importance, as with it images can be acquired without repositioning of the patient. Though a typical data-set of a whole-body MRI consists of 512x410x1400 voxel in average. Such data-sets cannot be evaluated completely in a contemporary and reliable manner without computational aid. This is especially the case if the volumes have to be compared to previous studies (follow-up studies). The project deals with the development of efficient methodologies for computer-aided interpretation of large medical data-sets as well as time series of those. By highlighting medical relevant regions in the image data, the physician is assisted and with that a higher effectiveness and cost efficiency in clinical routine can be achieved. The main topic of the project is the treatment of lymphom node disease patients, however a generalization of the proposed methods has to be possible.

In order to successfully finish the project, it requires a tight cooperation of computer scientists and physicians. The involved groups are the institute of pattern recognition (Informatik 5) of the Friedrich-Alexander University Erlangen-Nuremberg and the radiology and nuclear medicine of the Charité, Campus Benjamin-Franklin, Berlin. The workload for the institute of pattern recognition is the development of novel efficient methodologies to handle large medical data-sets. Their practicability in the clinical environment and their validity are evaluated by the involved physicians.

Conceptually the project can be separated in two disjunct approaches. First, lymph nodes are detected in MRI images of a single study. The second phase deals with the localization of the nodes in time sequences of whole-body MRI scans.

Detection of lymph nodes in a single MRI study

The detection of lymph nodes in a single MRI study bases on the evaluation of several weightings of MRI data-sets. The evaluated sequences are all studies used in clinical daily routine (e.g. T1-weighted, T2-weighted, FLAIR and TIRM sequences). A very important fact for the choice of the sequences was their acquisition time. First experiments show, that especially T1-weighted and TIRM sequences yield promising segmentation and localization results. In order to compare both data-sets, they have to be registered in an initial preprocessing step. As the acquisition time point does not vary very much between the two scans, the volumes are assumed to be already matched nearly perfectly. Nevertheless, to correct slight movements of the patient, a non-rigid registration of the images is performed. As the regarded data-sets are from the same modality but have different weightings, the distance measures used are originally designed for the alignment of multi-modal volumes (e.g. mutual information, normalized cross coefficient). However, due to the properties of non-rigid registration the plausibility of the matching has to be carefully watched, to guarantee a more simple segmentation problem. For the localization of the lymph nodes statistical methods are used only. This has two advantages: first, using these approaches usually leads to a probability of the detection success of the regarded structures, like lymph nodes for instance, correlating with the goals of the project. Second, statistical methods are more general and with that can be adapted to other localization problems more easily. For this purposes different classes of approaches are used. For further preprocessing of the data-sets, methods like probabilistic intensity adaption between the volumes and probabilistic subtraction imaging are utilized. For the detection of the nodes, the methods basically rely on the clustering of the data-sets by classifying all voxels utilizing fuzzy c-means or Markov random fields based methods for instance.

Detection of lymph nodes in time sequences

A further main topic of the project is the detection of lymph nodes in time series of MRI images. For follow up studies the evaluation of the required data-sets is very time consuming as more than one whole-body scan has to be treated in parallel. Hence the automatic localization step is desirable for the physician. As the single volumes are acquired at different time points, they have to be rigidly pre-registered so that they fit to each other. This is followed by a non-rigid registration to compensate non- linear deformations. The result of the alignment is a vector field that describes the deformations between the two data-sets according to a distance measure. With that the deformation field describes volume changes of evolving structures, like malign lesions for instance, too. There, growing structures are represented by mathematical sources while shrinking lesions can be detected as mathematical sinks. Together with the information gained from the previous localization step in MRI sequences of a single time point, this is used to detect changes of the lymph nodes between the acquired data-sets. Another possibility to gain information about local changes from registration results is to analyze the differences between the fixed and the transformed moving images. In order to achieve usable results the regularization of the non- rigid registration has to be very strict.

Presentation of the localization results

The main goal of the project is not to draw final medical conclusions but to support the physician in the evaluation of the acquired MRI data-sets. This is achieved by marking clinical relevant areas in the volumes. This is done by generating a probability map of the localization results. The map is presented as an overlay to the original data-sets. By choosing a threshold the physician is able to create confidence intervals and with that adjust the illustration to his needs.