Evaluation of Four Supervised Learning Schemes in White Matter Hyperintensities Segmentation in Absence or Mild Presence of Vascular Pathology

Abstract

We investigated the performance of four popular supervised learning algorithms in medical image analysis for white matter hyperintensities segmentation in brain MRI with mild or no vascular pathology. The algorithms evaluated in this study are support vector machine (SVM), random forest (RF), deep Boltzmann machine (DBM) and convolution encoder network (CEN). We compared these algorithms with two methods in the Lesion Segmentation Tool (LST) public toolbox which are lesion growth algorithm (LGA) and lesion prediction algorithm (LPA). We used a dataset comprised of 60 MRI data from 20 subjects from the ADNI database, each scanned once in three consecutive years. In this study, CEN produced the best Dice similarity coefficient (DSC): mean value 0.44. All algorithms struggled to produce good DSC due to the very small WMH burden (i.e., smaller than 1,500 \(\text {mm}^3\)). LST-LGA, LST-LPA, SVM, RF and DBM produced mean DSC scores ranging from 0.17 to 0.34.