Strategies for the Detection of Focal Lesions in MR Tomograms of the Human Head

Detecting pathological features in magnetic resonance imaging data sets of the human head is considered a non-trivial segmentation task. Segmentation approaches require prior knowledge about the lesion characteristics (e.g., their expected compartment, size and shape, their signal statistics in relation to the embedding tissue), and thus, are generally targeted for detecting a specific lesion type. We focus on approaches that are targeted to automatically detect (a) large unilateral lesions, (b) small multifocal lesions, and (c) diffuse white matter lesions. While a trained human observer still outperforms automatic approaches for lesion detection and discrimination, estimating the lesion size (in the case of large unilateral or diffuse lesions), count and position (in the case of multifocal lesions) is tedious or even impossible. The aim of this work is to provide quantitative descriptors for different lesion types. The three example algorithms discussed here incorporate prior knowledge in fundamentally different ways. These quantitative descriptors are useful measures for the statistical evaluation of a patient's clinical status.