Automatic identification of MS lesions based on local steering kernel features and sparse dictionary training

This paper proposes an automatic method for the identification of Multiple Sclerosis (MS) lesions in brain magnetic resonance (MR) images using Local Steering Kernel (LSK) features and a sparse dictionary discrimination approach. In the proposed method, wavelet transformation is initially used to extract salient points from the image. Next, LSK features of various sizes are extracted from patches around these salient points. A dictionary is then constructed and trained using Fisher’s sparse discrimination criterion. Expert-segmented MR images are used to train the dictionary. For each test image, a search is conducted within the dictionary according to the Fisher criterion, and classification is performed to identify MS lesions. The proposed method is fully automated and does not require any user initialization. Experimental results on the BrainWeb database show that the proposed method achieves a 5.99% improvement in Dice coefficient compared to state-of-the-art methods for MS lesions identification.