Am I misunderstanding Figure 1? It looks like even with a very aggressive high-pass filter (see HPF-50 in Figure 2), the model is still extremely accurate. That feels like a red flag to me - it makes me think there is some sort of leakage in the imaging that isn't being accounted for. It seems so implausible that you could tell race from the HPF-50 image that I'm worried I've completely misread the figure.