Validation of White Matter Hyperintensities automatic segmentation methods

Abstract

[en] This master’s thesis seeks to review and objectively evaluate the current white matter hyperintensities (WMH) automatic segmentation methods published journals. To this end, the methods have been systematically searched in scientific databases, and those meeting inclusion criteria have been evaluated. The evaluation has consisted in applying the method to detect WMH in our dataset of patients with bipolar disorder and healthy controls, in which an experienced neuroradiologist had manually coded all WMH. After the systematic search, we selected all available methods that were ready for use with standard MRI data by a standard user. Four methods met these criteria. We then applied these methods to detect WMH in our dataset, and compared the results with the neuroradiologist-based ground truth deriving several evaluation metrics. This master’s thesis also include a discussion section, in which we compare the results of our evaluations with the results of the WMH Segmentation Challenge held in 2017, which included substantially different datasets. The most relevant conclusion of this master’s thesis is that no method seems to be accurate enough for clinical implementation, although the low performance of the methods may be related to the differences between our data and the data that were used to train them. Besides, realizing the huge improvement made in the field during the last few years after the appearance of deep neural networks, we anticipate that a method with sufficient accuracy might be available soon. The codes used to obtain the results and graphs displayed in this project together with some guidelines to run them are available through PFM-WMH 1.