Souza, Roberto, et al. “An open, multi-vendor, multi-field-strength brain MR dataset and analysis of publicly available skull stripping methods agreement.” NeuroImage (2017).
From the past until now, during research, I wondered which preprocessing is best or which BET(brain extraction) selection is best for a brain pipeline. For solving this problem, I had to know a variety of pipeline knowledge in computation approach, then I needed to decide how to validate a hypothesis. However, these were beyond my capacity at that time. Therefore I just used freesurfer built-in BET or FSL(FMRIB Software Library) BET using default parameter.
After time goes by, when I found this paper, I had to be surprised.
This paper performed what I want to do and more than. First, the authors used a variety of age, gender-matched dataset from multi-vendor, multi-field-strength. However, although the age range is as matched as possible, there is a significant difference in statistically. So, the authors use age as a random effect in constructing a linear mixed model.
Second, the author performed most of the BET pipeline that exist in brain analysis for experiments. The author used a consensus method and presented supervised classification using linear regression as well as these BET pipelines. Especially, in this paper, the brain mask by supervised classifiers was called by ‘silver standard’. Then, they drew twelve segmentation brain mask called such as ‘gold standard’ using ITK tools. In statistical validation, It is regrettable that the authors did not verify reliability about ‘gold standard’.
The result of this paper show simple as how much brain tissue is left out, how much non-brain tissue is included, whether is affected by outliers or how much BET is similar. During experiments, the author only applied default parameter to each BET pipeline, individually.
Do you have considered which pipeline to use for BET in your problem? for better results? This paper shows that ANTs, BEaST is best performing and robust. If you want to find brain fissure well, you decide to use MBWSS. However, the method has a disadvantage that preserves the spinal cord.
In our lab, BET by FSL is widely used. Unfortunately, BET was not performed well than other pipelines and had high variance. So, it might be littel risky, sometime. However, the authors only used default parameter. If you checked the mask image visually, It’s OK. But, I forcely recommend ANTs if you want to perform BET well automatically.
In summary, BET is affected by vender or field-strength. This meaning is well-matched what I was intuitively thinking about.