Correction and methodological note. This poster is preserved as a record of an undergraduate project, but several claims need qualification. The age-bin counts shown on the poster sum to 1,206, whereas the text says that 1,113 participants had usable MRI scans; the former describes the larger behavioral cohort and should not be treated as the class distribution of the imaging model without filtering to the actual scan subset. In the displayed three-class counts, the largest class alone is $527/(247+527+418)\approx44.2\%$, so a quoted 33% “chance” baseline is appropriate only if the resampling procedure balanced the classes. Accuracy differences should also be accompanied by paired uncertainty or an equivalence test before being described as significant or as showing “no loss.”
The normalized shape ratio $T^2/S$ is dimensionless, as are the analogous isoperimetric ratios $P^2/A$ in two dimensions and $S^3/V^2$ in three dimensions. However, the poster's universal-approximation remark is incorrect: a continuous rational function on a compact domain that avoids its poles lies within the usual class of functions that a sufficiently wide feed-forward network can approximate uniformly. The poster and linked two-page research snapshot do not contain enough implementation and resampling detail to reproduce the reported classification results, so those results should be read as preliminary.
