Multimodal spatial-based segmentation framework for white matter lesions in multi-sequence magnetic resonance images

Abstract

Abstract Objective Multi-sequence magnetic resonance (MR) imaging is a frequently used method for characterising and quantifying white matter (WM) lesions in the human brain. The number and size of lesions are commonly determined to assess the diseases in clinical settings. Accurate WM lesion segmentation is very important for disease diagnosis and progression surveillance. The goal of this paper is to present an approach for improving WM lesion segmentation accuracy. Methods In this paper, we propose a novel method integrating the multi-sequence and spatial information in a Bayesian framework for WM lesion detection from multi-sequence human brain magnetic resonance images (MRIs). The entire framework is based on a three-step approach: First, a multinomial logistic regression (MLR) algorithm is used to assess the conditional probability distributions of intensities in WM lesions and brain tissues from training data. Second, the spatial information previously given by a Markov random field (MRF) prior is integrated with multimodal information in the Bayesian framework to strengthen the spatial constraint. This step is especially effective when WM lesions have intensity values similar to those of other brain tissues. Finally, a post-processing step based on biological knowledge is used to remove some false positives. Results Our method is validated using two datasets. The experimental results show that our algorithm agrees well with manual expert labelling and indicate that our multimodal spatial-based method offers a significant advantage over other approaches. Conclusions A three-step approach for combining multimodal and spatial information is proposed for WM lesion segmentation. The advantages of this approach are discussed, and a practical application to two datasets is presented.