An Accurate and Efficient Segmentation Method of White Matter Hyper-intensity using Deep LearningDownload PDF

09 Feb 2021 (modified: 16 May 2023)Submitted to MIDL 2021Readers: Everyone
Keywords: Semantic Segmentation, Deep Learning, White Matter Hyperintensities, Connected Components Algorithm, Medical Imaging
TL;DR: We developed PVWMH/DWMH segmentation model using deep learning and connected components algorithm based on a "continuity to ventricle" rule which is widely used as clinical diagnostic criteria
Abstract: Unlike the rich study of automatic segmentation of white matter hyperintensity (WMH), only a few divide WMH into Deep WMHs (DWMH) and Periventricular WMHs (PVWMH). Distinguishing the two is critical because of their different clinical implications. In this work, we propose a novel WMH segmentation method which segments both DWMH and PVWMH from T2-FLAIR MR images. The proposed method operates in two steps: WMH separating segmentation and subdivision. In the first step, convolutional neural networks (CNNs) are trained to separate PVWMH from DWMH segmentation, and both masks were combined into one whole WMH mask in the first stage. The whole WMH mask is then divided into DWMH and PVWMH via lateral ventricle segmentation and post-processed through a connected components algorithm. This post-processing adheres to the "continuity to ventricle" clinical criterion for dividing PVWMH and DWMH based on the segmented lateral ventricles. The proposed method not only reduces false negatives, achieving a Dice coefficient score of 0.83, but also partitions WMH into PVWMH and DWMH, enabling a more fine-grained diagnosis.
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.
Paper Type: methodological development
Primary Subject Area: Segmentation
Secondary Subject Area: Application: Radiology
6 Replies

Loading