Fig. 3

Analysis of prediction values from submodules to understand the model’s decision. a Acquiring prediction values from submodules along the 3 axes. b The Aβ-positive and Aβ-negative PET scan participants are divided based on the final classification, and a two-sample t test is performed to determine substantial slice numbers. c) Points are plotted for each slice number in the MNI space and correlated labels are extracted from the AAL atlas. In the example, if a significant slice number is x = 46, 51, y = 34, 64, 66, and z = 37, 45, a total of 12 points are plotted in the MNI space. However, actual significant points are (46,34,37), (51,54,45), and (51,66,45). d) To exclude unrelated points, ROIs in the AAL atlas with more than 50 points are used for comparison with the result of the voxel-wise analysis. Abbreviation: Aβ = β-amyloid; MNI = Montreal Neurological Institute; AAL = Automated Anatomical Labeling; ROI = region of interest