Comparison between tools for automatic segmentation of white matter hyperintensities of presumed vascular origin in aging: which one, how and why to choose the most suitable for your purpose?

Abstract

AbstractCerebrovascular damage consequent to small vessel disease (SVD) is a common companion of healthy and pathological aging. Neuroimaging signatures of SVD are present in virtually all people older than 60 years, and their prevalence increases with age. According to the STRIVE criteria, white matter hyperintensities (WMH) have been associated with different clinical symptoms and constitute a good clinical proxy for SVD. This is important as WMH can be directly assessed via MRI. Currently, the most widely used method to detect and assess the severity of HWH are clinical scales based on visual assessment, but these scales do not offer real quantitative information, making it difficult to assess progression. Quantitative information can be approached through the manual segmentation of physicians, but this process is extremely time-consuming and presents a high inter and intra evaluators variability, which makes its application in routine protocols unfeasible. Therefore, it is imperative to facilitate the use of automatic protocols capable of providing WMH load measurements that are as accurate as possible to those obtained by manual segmentation. In this study, we aim to identify the most accurate software for WMH segmentation, providing not only methodological insights but also usability knowledge that would enlighten the tradeoff between clinical accuracy and real-world implementation. The data set consisted of T1 and Flair images of 45 cognitively healthy older participants (mean age 71 ± 5). The study analyzed WMH segmentations obtained with clinician manual segmentation and four tools included in three of the most widely used neuroimaging toolkits: 1) Lesion Prediction Algorithm (LPA) and Lesion Growth Algorithm (LGA) of the lesion segmentation tool (LST); 2) sequence adaptive multimodal segmentation (SAMSEG); and 3) the brain intensity anomalies classification algorithm (BIANCA). The analysis evaluated the correlations with the Fazekas clinical scale, the influence of the WMH load and evaluated the performance at the individual lesion level. The results showed that the supervised methods (LST-LPA and BIANCA) performed better in all the analyzes and were the only ones capable of consistently capturing small lesions (<26 mm3). However, these tools lose performance when applied to new data. Considering the results (accuracy and ease of use), we concluded that, in the general case, the combined use of the two FSL tools emerged as the best option. We confirmed this conclusion by evaluating WMH segmentation in a dataset of 500 older individuals and found that the LPA results, using LGA to control for divergences, offered valuable and actionable clinical information, both to help clinicians make treatment decisions treatment and to monitor pathological progression.