The trustworthiness problem has two distinct faces, and the team led by Majid Ramedani and Martin Dyrba attacked both. The first is demographic confounding. Aging is the single greatest risk factor for Alzheimer’s disease, and the structural overlap between normal aging and pathology means a disease classifier may inadvertently rely on aging patterns rather than disease-specific atrophy. Brain age prediction models offer telling evidence of this: they typically estimate older ages for patients with neurodegenerative disease, a signature that the same confounding signal has leaked into the learned representations. The second face is unreliable probability estimates. The softmax layer in a convolutional neural network transforms raw output scores into a probability distribution over classes, but those disease likelihood scores indicate only the model’s relative confidence — not true statistical probabilities — and they come without confidence intervals. A model that says a patient has a 0.8 probability of Alzheimer’s is making a claim about its own internal scores, not about observed disease frequencies.

To test their solutions rigorously, the researchers built all their models on a shared 3D convolutional neural network architecture, ensuring that any performance differences could be attributed to the debiasing strategy rather than to architectural advantages. The baseline CNN consists of four convolutional blocks with 3×3×3 kernels, batch normalization, ReLU activation, and 2×2×2 max-pooling, followed by a bottleneck module that compresses whole-brain voxel-wise gray matter probability maps into a 4,096-dimensional feature vector. The inputs were whole-brain, voxel-wise gray matter maps derived from T1-weighted MRI scans, spatially normalized to MNI space and processed through a single standardized pipeline applied identically to all three cohorts. The models operated directly on whole-brain voxel-wise information rather than on predefined regional volumes, letting the networks find their own features in the raw gray matter signal.

The first debiasing approach, called CF-Net, is an adversarial confounder-free network built on a min–max optimization game. A feature extractor generates a feature vector that a classifier uses to predict diagnosis, while a separate confounder predictor attempts to predict the patient’s age from the very same features. The adversarial loss is defined using the squared Pearson correlation between true and predicted confounders, and minimizing it drives that correlation toward zero, encouraging statistical independence between the learned features and age. Training proceeds in three phases: the encoder and classifier train jointly first, then the confounder predictor activates while the others freeze, and finally a joint loop lets the encoder and confounder predictor interact to refine the representation. Learning rates follow an 8:1 ratio to balance the capacities of the 6.3-million-parameter classifier and the 0.8-million-parameter confounder predictor.

The second approach, PMDN — penalty-based metadata normalization — takes a fundamentally different route. Instead of playing an adversarial game, it treats the confounding effect as a general linear model, modeling the feature vector as f minus Mβ equals r, where M contains metadata, β are regression coefficients, and r is the residual, confounder-free feature vector. Because deep learning optimizes batch-wise, naive least-squares estimates of β would vary across batches and introduce instability. PMDN solves this by formulating a differentiable, bilevel optimization problem that jointly updates the network weights and the deconfounding parameters, with a penalty term that suppresses feature correlation with metadata. Alternating updates of the weights and β ensure stable convergence while the network learns task-relevant, residualized features. At inference, only the confounding components are subtracted from the 4,096-dimensional encoder features.

The results, tested on external cohorts after training on ADNI data, tell a striking story about stability. PMDN achieved the highest balanced accuracy at 0.71 with a standard deviation below 0.02 — roughly ten times lower than the baseline’s variability of about 0.11 and five times lower than CF-Net’s 0.06. All models posted comparable AUC values around 0.82, meaning their ability to discriminate diseased from healthy brains was essentially identical. But PMDN delivered the lowest Brier score at 0.22, against 0.32 for the baseline, indicating far more reliable probability estimates. The age bias quantification was even more revealing. On an age-balanced test set, the baseline model showed a consistent positive correlation between chronological age and predicted Alzheimer’s logits in cognitively normal subjects (r = 0.16), meaning older healthy people were systematically assigned higher disease risk. CF-Net barely moved that needle (r = 0.15) with greater variability, but PMDN reduced it to near zero (r = −0.07).

The mild cognitive impairment subgroup — the clinically critical category where age-related atrophy and pathological changes overlap most — showed the clearest separation between strategies. The baseline model exhibited a substantial age-logit correlation of r = 0.19, and CF-Net was actually worse at r = 0.20, suggesting adversarial training was ineffective at mitigating age bias in this population. PMDN achieved a near-zero correlation of r = 0.06, reflecting successful removal of age-confounded variance exactly where it matters most for early detection. In the diagnosed Alzheimer’s group, however, age-related associations persisted across all three models, with correlations around r = 0.15 unchanged by either debiasing method. The researchers attribute this to the greater overlap between age-related brain changes and disease-related neurodegeneration in established pathology, which makes it genuinely harder to disentangle the two signals. Even so, PMDN showed lower variability in the age-logit association for the Alzheimer’s group, indicating more consistent behavior across evaluation folds.

The second half of the study tackled the calibration problem with four post hoc methods applied to the CNN’s softmax outputs: two parametric approaches (logistic regression and probit regression) and two nonparametric ones (isotonic regression and Bayesian binning into quantiles, or BBQ). Logistic regression applies Bayes’ theorem under Gaussian class-conditional assumptions, reducing to an affine transformation of the raw scores, while probit regression substitutes the inverse normal link function. Isotonic regression fits a nondecreasing, piecewise-constant function using the pool-adjacent-violators algorithm, achieving an optimal linear-time solution without assuming any parametric form. BBQ divides outputs into bins based on empirical quantiles and estimates probabilities within each bin using a Bayesian beta-distribution approach, averaging over models with varying bin counts. The uncalibrated CNN outputs showed the largest calibration errors, with an expected calibration error of 0.28 and a maximum calibration error of 0.64 — meaning the model’s stated confidence could deviate from observed accuracy by nearly two-thirds at its worst.

Calibration transformed those numbers dramatically. Parametric methods produced moderate improvements, cutting the expected calibration error to roughly 0.13 and the maximum to about 0.29 with smooth, constrained curves. The nonparametric approaches did better still: isotonic regression and BBQ achieved expected calibration errors of 0.07–0.08 and maximum errors of 0.19–0.20, with isotonic also posting the lowest log-loss (0.50) and Brier score (0.16) while preserving the same F1-score of 0.73. Classification performance remained largely stable across all methods, confirming that calibration improves the reliability of predicted probabilities rather than the model’s discriminative ability. The researchers note that monotonic approaches produced smooth and reliable calibration curves, whereas the nonmonotonic BBQ method introduced irregularities — a reminder that flexibility in calibration comes at the cost of potential instability, and that evaluation should combine metrics like ECE and MCE with visual inspection of reliability plots.

The deeper lesson of the study is that deep learning models achieving similar classification performance can differ dramatically in calibration, stability, and fairness — dimensions that headline accuracy figures never capture. The contrast between the two debiasing strategies stems from their underlying optimization: PMDN applies a deterministic GLM-based correction that produces stable, consistent feature representations, while adversarial training relies on a min–max game with competing objectives that may converge incompletely, leaving residual confounding and explaining CF-Net’s variable performance. The findings carry real clinical weight for mild cognitive impairment populations, where models that lean on age as a proxy risk overestimating disease probability in older individuals. The team acknowledges limitations: age was the only confounder explicitly addressed, while sex, scanner variability, and site effects may also shape predictions; dataset size remains limited relative to the dimensionality of neuroimaging data; and adversarial training might reach its full potential with longer or more carefully tuned optimization. Future work will extend the approach to multiconfounder settings, larger and more diverse cohorts, and anatomically informed regional inputs such as hippocampal measures, which may encourage reliance on clinically established disease markers while reducing the influence of widespread age-related structural change.

Subject of Research: Artificial Intelligence

Article Title: Improving the fairness and robustness of deep learning classification models through bias mitigation and post hoc calibration

Article References: Ramedani, M., Terli, J. C., Martensen, O. H., Strauß, T., Buerger, K., Dechent, P., Düzel, E., Fliessbach, K., Gemenetzi, M., Glanz, W., Hellmann-Regen, J., Hetzer, S., Incesoy, E. I., Janowitz, D., Kronmüller, M. T., Laske, C., Lüsebrink, F., Perneczky, R., Peters, O., … Dyrba, M. (2026). Improving the fairness and robustness of deep learning classification models through bias mitigation and post hoc calibration. International Journal of Computer Assisted Radiology and Surgery. https://doi.org/10.1007/s11548-026-03805-x

Image Credits: AI Generated

DOI: 10.1007/s11548-026-03805-x

Keywords: deep learning, Alzheimer's disease, machine learning, medical imaging, MRI, algorithmic fairness, calibration, neurodegenerative disease, Improving, fairness, robustness, deep