Description
Deep learning classifiers can diagnose Alzheimer disease using structural MRI with high accuracy. However, they cannot explain to a clinician what anatomical changes caused a prediction or what minimal changes can flip the decision. This hinders their clinical implementation. Counterfactual explanations answer both questions: by creating an input that is minimally different and crosses the decision boundary. They reveal the structural properties that the model considers to be decisive. We present a Brain Latent Progression Extension (BrLP-X) framework. A 3D variational autoencoder is used to reduce whole-brain structural MRI to a latent space, followed by a conditional diffusion model that, with the help of diagnostic labels, synthesizes counterfactual brain volumes. A structural causal model imposes the known sequence of AD neurodegeneration, beginning with early entorhinal thinning, through hippocampal atrophy to extensive cortical engagement, by obscuring areas that are not yet diseased at the target disease stage. We use Prototypical Networks (ProtoNets) as a classifier. Even though ProtoNets are said to be interpretable at the model level, local, instance-level explanations are still useful to individual predictions, and our counterfactuals offer such explanations. The training is based on 1,932 structural MRI volumes in ADNI that are used in four diagnostic groups (CN, sMCI, pMCI, AD). Our generated counterfactuals have high visual fidelity (SSIM 0.9979–0.9980, PSNR 57.74–58.36 dB) with latent FID ≤ 0.24, and with sparse and anatomically localizedregion-level changes (6.7–12.3% sparsity, L0 ≈ 150). Compared to gradient, Wachter, and DICE baselines, the SCM-constrained variant decreases the distance to the target in the form of L2 by 23–58%, improves saliency Dice with known AD atrophy regions to 0.37, and lowers Mahalanobis distance to the target class by 15–33%. Bootstrap 95% confidence intervals and an observed zero SCM-violation rate across all scenarios confirm that the causal graph effectively shapes and constrains generated counterfactuals.| Period | 19 Jun 2026 |
|---|---|
| Event title | Yorkshire Innovation in Science and Engineering Conference 2026 |
| Event type | Conference |
| Location | Huddersfield, United KingdomShow on map |
| Degree of Recognition | National |