Analysis of a 3D brain MRI sex classifier via Approximate Inverse Model Explanations

In this study, we have analyzed the prediction rationale of a deep learning model for sex classification from 3D brain MRI using Approximate Inverse Model Explanations (AIME). A 3D DenseNet121 classifier has been trained on 566 T1-weighted IXI scans. The model has achieved 98.2% accuracy on a 114-case validation set. Global importance has shown a sign-reversal pattern between classes: peripheral regions contribute to Male prediction, whereas central regions contribute to Female prediction. Local importance has been consistent with this pattern and has highlighted strong peripheral reliance in misclassified cases. Controlled experiments (skull-stripped retraining, masking sensitivity, and age-matched analysis) have indicated substantial dependence on extra-brain information. Cross-validation and leave-one-site-out evaluation have supported the robustness of these findings.