Study On Detection Algorithm Of White Matter Lesion In Multi-modality MRI

Leukoaraiosis (LA) is a white matter lesion and it appears as multiple patchy hyperintense inperiventricular and subcortical areas of T2FLAIR image. LA is often found in aged individuals, andthe pathogenesis of these white matter abnormalities is still unclear. Many studies show that it maybe associated with stroke, Parkinsonism, mild cognitive impairment, Alzheimer’s disease, and evenhas been linked to mood disorders. At the same time, in the area of ESWAN imaging modecorresponding to the zone of LA in T2FLAIR imaging mode, deep medullary veins (DMVs) arerelevant visible. The relationship between them has been explored by medical community in recentyears, and the researchers are trying to analysis the reasons by medical imaging. But the DMVs areshowed with characteristics of small diameter, imaging blur and difficult to distinguish. Cliniciansusually determine the level of LA and DMVs by subjective experience or simple grid measurement.These subjective evaluations cannot provide an accurate argument for scientific research. Thereforea reliable computer aided method is necessary for helping doctors accurately segment the LAlesions at T2FLAIR mode and extract the DMVs at ESWAN mode.Against such a demand, we studied a novel approach for LA segmentation and DMVsextraction based on the imaging characteristics under different imaging mode of these two. Theproposed novel automatic detection method can be divided into three steps, the first step is removalof non-brain tissue, the second step is LA segmentation in T2FLAIR mode, finally is DMVs trunkextraction in corresponding ESWAN mode. Skull stripping pretreatment combines Otsu thresholdmethod and mathematical morphology processing, and this method can strip skull effectively. Wemake a control experiment with the popular BSE skull stripping method, the results show that theproposed method is more accuracy than BSE method from Jaccard similarity and Dice similarity.LA segmentation finds out the initial contour by Otsu method first, and then evolution the contourby a modified C-V model. Experiments prove that this method performed slightly better on thesegmentation results and performed better on the segmentation time compared with Li’s non-reinitialization level set method. At last we design a filter according to the relationship betweeneigenvalues of Hessian matrix and grayscale directions of DMVs. The filter preliminary screenseffective trunks and then combined with morphology post-processing, including connecting disconnected trunks and removing artifacts. The experimental results show that the effect of thismethod are very close to manual extraction.