Multiple sclerosis (MS) lesion segmentation is crucial for evaluating disease burden, determining disease progression and measuring the impact of new clinical treatments. MS lesions can vary in size, location and intensity, making automatic segmentation challenging. In this paper, we propose a new supervised method to segment MS lesions from 3D magnetic resonance (MR) images using non-local means (NLM). The method uses a multi-channel and rotation-invariant distance measure to account for the diversity of MS lesions. The proposed segmentation method, rotation-invariant multi-contrast non-local means segmentation (RMNMS), captures the MS lesion spatial distribution and can accurately and robustly identify lesions regardless of their orientation, shape or size. An internal validation on a large clinical magnetic resonance imaging (MRI) dataset of MS patients demonstrated a good similarity measure result (Dice similarity?=?60.1% and sensitivity?=?75.4%), a strong correlation between expert and automatic lesion load volumes (R 2 ?=?0.91), and a strong ability to detect lesions of different sizes and in varying spatial locations (lesion detection rate?=?79.8%). On the independent MS Grand Challenge (MSGC) dataset validation, our method provided competitive results with state-of-the-art supervised and unsupervised methods. Qualitative visual and quantitative voxel- and lesion-wise evaluations demonstrated the accuracy of RMNMS method. Graphical abstract Display Omitted Highlights ? We propose a new multi-channel MS lesion segmentation technique. ? We adapt for lesion segmentation the non-local means operator to account for multi-contrast and rotation-invariant distance. ? The proposed method presents highly competitive results compared to state-of-the-art methods. ? The proposed method provides segmentation quality near inter-rater variability for MS lesion segmentation. ? Our non-local approach is able to detect structures that vary in size, shape and location such as MS lesions.
展开▼
