Accurate white matter lesion segmentation by k nearest neighbor classification with tissue type priors (kNN-TTPs)

• Intensity normalization has a large influence on lesion segmentation performance.
• Inclusion of tissue type priors as features increases segmentation performance.
• Best performance was achieved using variance scaling and tissue type priors.

IntroductionThe segmentation and volumetric quantification of white matter (WM) lesions play an important role in monitoring and studying neurological diseases such as multiple sclerosis (MS) or cerebrovascular disease. This is often interactively done using 2D magnetic resonance images. Recent developments in acquisition techniques allow for 3D imaging with much thinner sections, but the large number of images per subject makes manual lesion outlining infeasible. This warrants the need for a reliable automated approach. Here we aimed to improve k nearest neighbor (kNN) classification of WM lesions by optimizing intensity normalization and using spatial tissue type priors (TTPs).MethodsThe kNN-TTP method used kNN classification with 3.0 T 3DFLAIR and 3DT1 intensities as well as MNI-normalized spatial coordinates as features. Additionally, TTPs were computed by nonlinear registration of data from healthy controls. Intensity features were normalized using variance scaling, robust range normalization or histogram matching. The algorithm was then trained and evaluated using a leave-one-out experiment among 20 patients with MS against a reference segmentation that was created completely manually. The performance of each normalization method was evaluated both with and without TTPs in the feature set. Volumetric agreement was evaluated using intra-class coefficient (ICC), and voxelwise spatial agreement was evaluated using Dice similarity index (SI). Finally, the robustness of the method across different scanners and patient populations was evaluated using an independent sample of elderly subjects with hypertension.ResultsThe intensity normalization method had a large influence on the segmentation performance, with average SI values ranging from 0.66 to 0.72 when no TTPs were used. Independent of the normalization method, the inclusion of TTPs as features increased performance particularly by reducing the lesion detection error. Best performance was achieved using variance scaled intensity features and including TTPs in the feature set: this yielded ICC = 0.93 and average SI = 0.75 ± 0.08. Validation of the method in an independent sample of elderly subjects with hypertension, yielded even higher ICC = 0.96 and SI = 0.84 ± 0.14.ConclusionAdding TTPs increases the performance of kNN based MS lesion segmentation methods. Best performance was achieved using variance scaling for intensity normalization and including TTPs in the feature set, showing excellent agreement with the reference segmentations across a wide range of lesion severity, irrespective of the scanner used or the pathological substrate of the lesions.