International Journal of Computer Assisted Radiology and SurgerySource publication:
Loading content…
On structural MRI from three independent cohorts, Alzheimer's disease classifiers use age as a shortcut: older patients receive higher predicted risk. Residualizing age linearly in feature space (PMDN) brings the age-prediction correlation near zero in cognitively normal and mild cognitive impairment groups, with balanced accuracy 0.71 and Brier score about 0.22, and far less fold-to-fold spread than the adversarial CF-Net. Post hoc calibration further lowers probability error, with nonparametric isotonic regression and Bayesian binning performing best.
Age bias can be subtracted explicitly at the network feature level rather than through adversarial training. The undebiased baseline shows a positive association between age and predicted logit in cognitively normal subjects (r=0.16±0.03), assigning higher risk to older individuals. Trained on ADNI and tested on the external AIBL and DELCODE cohorts, PMDN lowers the cognitively normal correlation to r=-0.07±0.06 and the mild cognitive impairment correlation from 0.19±0.07 at baseline to 0.06±0.03.
The debiasing strategies differ mainly in stability and probability reliability, not in discrimination. AUC is comparable across models, at 0.82±0.02 for CF-Net and 0.82±0.01 for PMDN. PMDN reaches balanced accuracy 0.71±0.01 with fold standard deviation under 0.02, about ten times lower than the baseline (0.66±0.11) and five times lower than CF-Net (0.70±0.06); its Brier score is 0.22±0.01 against 0.32±0.16 for the baseline.
Age effects are treated as covariates in a generalized linear model, with network weights and deconfounding coefficients updated alternately during training. The original metadata normalization solved coefficients by least squares over the whole training set, so batch-wise training let them drift between batches. PMDN regresses 4096-dimensional encoder features onto three metadata terms through a differentiable bilevel optimization with penalty weight 0.9, subtracting only the confounding component at inference.
Post hoc calibration improves the probability estimates themselves while classification performance stays largely unchanged. Uncalibrated CNN outputs carry the largest calibration error, with expected calibration error 0.28 and maximum calibration error 0.64. Logistic and probit regression cut expected calibration error to about 0.13 and maximum calibration error to about 0.29; isotonic regression and Bayesian binning reach about 0.07–0.08 and 0.19–0.20 while accuracy near 0.77 and F1 near 0.73 are preserved.
Results cover a three-cohort structural MRI setting with ADNI for training and AIBL plus DELCODE for external testing, whole-brain voxel-wise gray matter probability maps as input, and age as the only explicitly handled confounder. The intended readers are researchers and engineers concerned with fairness and probability calibration in medical imaging models. The source proposes extending the approach to multiconfounder settings, larger and more diverse cohorts, and disease-focused regional inputs such as hippocampal or medial temporal measures.
Debiasing is uneven across diagnostic groups: age-prediction associations persist in the AD group for all three models, at 0.15±0.07 for the baseline, 0.17±0.07 for CF-Net and 0.15±0.04 for PMDN, with PMDN only showing lower fold-to-fold variability. PMDN trades larger absolute prediction errors for better calibration, and paired t-tests show its prediction error significantly above the baseline (mean difference -0.038, p<0.001). The source also notes that dataset size remains limited relative to the high dimensionality of neuroimaging data and that adversarial training may benefit from longer or more carefully tuned optimization.