Abstract
We propose a fast, learning-based algorithm for segmenting white matter (WM) lesions for magnetic
resonance (MR) brain images. The inputs to the algorithm are T1, T2, and FLAIR images. Unlike
most of the previously reported learning-based algorithms, which treat expert labeled lesion map
as ground truth in the training step, the proposed algorithm only requires the user to provide a few
regions of interest (ROI's) containing lesions. An unsupervised clustering algorithm is applied to
segment these ROI's into areas. Based on the assumption that lesion voxels have higher intensity
on FLAIR image, areas corresponding to lesions are identified and their probability distributions
in T1, T2, and FLAIR images are computed. The lesion segmentation in 3D is done by using
the probability distributions to generate a confidence map of lesion and applying a graph based
segmentation algorithm to label lesion voxels. The initial lesion label is used to further refine the
probability distribution estimation for the final lesion segmentation. The advantages of the proposed
algorithm are: 1. By using the weak labels, we reduced the dependency of the segmentation
performance on the expert discrimination of lesion voxels in the training samples; 2. The training
can be done using labels generated by users with only general knowledge of brain anatomy and
image characteristics of WM lesion, instead of these carefully labeled by experienced radiologists;
3. The algorithm is fast enough to make interactive segmentation possible. We test the algorithm
on nine ACCORD-MIND MRI datasets. Experimental results show that our algorithm agrees
well with expert labels and outperforms a support vector machine based WM lesion segmentation
algorithm.
