Measuring the prediction performance of Alzheimer’s disease using MR images with a deep learning model supported by a cross-replication strategy

This study evaluates the performance of a deep learning framework supported by a cross-replication strategy for predicting Alzheimer’s disease (AD) from structural magnetic resonance imaging (MRI). EfficientNetV2-B0 was selected due to its favorable accuracy-efficiency trade-off. The workflow consisted of two stages: (i) clustering-based relabeling of the full dataset into five clinically meaningful categories, and (ii) training a classifier on the relabeled data. To assess the stability of the proposed approach, the model was trained across multiple random initializations on a fixed train/validation/test split. Class-wise Average Precision, macro- and micro-averaged Precision-Recall Area Under the Curve (PR–AUC) and Receiver Operating Characteristic Area Under the Curve (ROC–AUC), and their 95% confidence intervals were reported using bootstrap resampling. The cross-replication strategy yielded improved stability across initializations, with a mean test accuracy of 0.95 compared with 0.94 for the single-run baseline, along with consistently higher PR–AUC and ROC–AUC values. These findings suggest that cross-replication enhances the reliability of AD stage prediction by reducing performance variability due to stochastic initialization, although further evaluation with alternative data partitions or external validation cohorts is warranted.