Using deep learning to quantify total scores of cerebral small vessel disease (CSVD) in a stroke cohort

Abstract

Cerebral small vessels include the arterioles, capillaries, and venules in the brain, which are essential for controlling cerebral blood flow and maintaining brain homeostasis. Diseases related to these vessels are defined as cerebral small vessel disease (CSVD). CSVD has multiple clinical presentations, including cognitive i impairment, gait disturbance, and acute ischemic stroke or transient ischemic attack. MRI features of CSVD include recent small subcortical infarct (RSSI), lacunes, enlarged perivascular space (EPVS), white matter hyperintensities (WMH), and cerebral microbleeds (CMB). To evaluate the total CSVD load in patients, a score is applied to assess each subtype of CSVD. However, it is very tedious and time-consuming to label these by hand. This thesis proposed an auto pipeline for the computer-assisted detection of CSVD using deep learning on a large dataset of local stroke patients. A total number of 974 subjects—all of whom had been clinically diagnosed with transient ischemic attack or ischemic stroke—were recruited in this study. An external testing cohort comprising 48 stroke patients was also collected for this study. These patients all underwent scanning at the local MRI unit and all patients were well-informed about this research and provided signed consent. All the MRI data were processed under the standards of the local MRI unit. The first chapter of this thesis provides more details about this project. In chapter 2 to chapter 6, the detection of each subtype of CSVD is discussed, covering CMB, EPVS, WMH, lacunes, and RSSI, correspondingly. The detection of WMH and CMB has been widely discussed in the literature, but we train the model to provide more clinical information rather than just focusing on the sensitivity. The detection of EPVS and lacunes has been less discussed previously, and we explore the application further. As for RSSI, few papers have reported on this subtype, and our results are discussed in chapter 6. A comparison of the total CSVD score generated by our proposed pipeline and the one labeled by clinical doctors is discussed in chapter 7. The two-tailed sample t-test and the Bland-Altman map were used for statistical analysis. ii

This thesis represents the first attempt to employ deep learning methods for the automated detection of total CSVD scores in a comparatively large cohort of stroke patients. The dataset utilized encompasses all subtypes of CSVD and was meticulously labeled by experienced clinical practitioners. This thesis provides a thorough and detailed application of deep learning detection on medical images and potentially opens avenues for robust applications in the field of AI-medicine. Abstract word counts: 424 words