A Novel Hierarchical Level Set with AR-boost for White Matter Lesion Segmentation in Diabetes

Hierarchical as well as coupled level sets are widely used for multilevel image segmentation. However, these tools are successful if the number of levels of an image are known and a careful choice of initialization is performed. We intend a novel hierarchical level set (HLS) followed by an Adaptive Regularized Boosting (AR-Boost) for automatic White Matter Lesion (WML) segmentation from Magnetic Resonance Images. HLS does not need to know the number of levels in an image and HLS is computationally less expensive and more initialization independent than coupled level-setssince HLS doesn't generate redundant regions. We employan energy functional that minimizes the negative logarithm of variances between the two partitions created by the level set function. HLS uses a level set to partition the image into a number of segments, then applies the level set on all the segments separately to create more segments and the process continues iteratively until all the segments become a nearly homogeneous region (low intensity variance). Then AR-Boost classifies the segments into WML and non-WML classes. The proposed loss function for AR-boost enforces more weight on misclassified samples at each iteration than Adaboost to classify correctly in the next iteration and consequently leads to early convergence. Unlike Adaboost, the user can select optimal weights through cross-validation. Experimental results demonstrate that the proposed method outperforms state-of-the-art automated white matter lesion segmentation techniques.