Extracting evolving pathologies via spectral clustering.

Abstract

A bottleneck in the analysis of longitudinal MR scans with white matter brain lesions is the temporally consistent segmentation of the pathology. We identify pathologies in 3D+t(ime) within a spectral graph clustering framework. Our clustering approach simultaneously segments and tracks the evolving lesions by identifying characteristic image patterns at each time-point and voxel correspondences across time-points. For each 3D image, our method constructs a graph where weights between nodes capture the likeliness of two voxels belonging to the same region. Based on these weights, we then establish rough correspondences between graph nodes at different time-points along estimated pathology evolution directions. We combine the graphs by aligning the weights to a reference time-point, thus integrating temporal information across the 3D images, and formulate the 3D+t segmentation problem as a binary partitioning of this graph. The resulting segmentation is very robust to local intensity fluctuations and yields better results than segmentations generated for each time-point.