AUTHOR=Shan Wei , Duan Yunyun , Zheng Yu , Wu Zhenzhou , Chan Shang Wei , Wang Qun , Gao Peiyi , Liu Yaou , He Kunlun , Wang Yongjun 

TITLE=Segmentation of Cerebral Small Vessel Diseases-White Matter Hyperintensities Based on a Deep Learning System

JOURNAL=Frontiers in Medicine

VOLUME=Volume 8 - 2021

YEAR=2021

URL=https://www.frontiersin.org/journals/medicine/articles/10.3389/fmed.2021.681183

DOI=10.3389/fmed.2021.681183

ISSN=2296-858X

ABSTRACT=Objective:
Clinical diagnoses based on a well-established system for automatic segmentation of white matter hyperintensities (WHMs) are useful and efficient in practice.
Method:
This is a retrospective study based on MRI-FLAIR data sets from 1156 patients (median age, 54 years; 653 males, 503 females) obtained between September 2018 and September 2019. All patients had been diagnosed with cerebral small vessel disease (CSVD) with clear MRI-FLAIR WMH characteristics. The patients were divided into training, validation, and testing cohorts of 870, 126, and 160 patients, respectively. All data for the model training were labeled layer by layer according to the consensus from two neuroradiologists with 15 years of work experience. After that, a two-dimensional convolutional neural network (CNN) was applied to the MRI data from the (870+126)-patient training and validation MRI cohort to construct a deep layer system (DLS), which was tested with the 160-patient, independent MRI cohort. Next, the DLS tool was used as a segmentation program for 90 patients from 3 different medical centers, independed from previous 1156 patients. These results were evaluated by 3 neuroradiologists and given an output analysis score, which was divided into 4 grades (accepted, minor revision, major revision, not accepted). The inter-neuroradiologists agreement rate, assessed by the Kendall-W test, was used for quality analysis.
Results:
In the detection and segmentation of the WMHs of 90 patients from 3 different medical centers, the DLS achieved a Dice coefficient of 0.72. In the clinical evaluations of the independent data from the three centers, the neuroradiologists reported that more than 95% of the results could be directly accepted or with minor revisions.
Conclusion:
The results demonstrate that the automatic detection and segmentation of WMHs is feasible. The proposed well-trained DLS system could be a trusted tool for the segmentation and detection of WMH lesions.