To explore deep learning-based recognition and classification methods to address the cumbersome and time-consuming issues of traditional Alzheimer's disease (AD) detection and assessment, thereby improving diagnostic and predictive speed and accuracy.

MRI data were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI). During preprocessing, MIPAV medical image processing software and 3D Slicer were used for AC-PC alignment, spatial normalization, skull stripping, and bias correction. The preprocessed data were converted into 5D tensors and input into a voxel residual network (VoxResNet) model in PyTorch for training. The training process employed L2 regularization, stochastic gradient descent (SGD) with momentum, and cross-entropy loss for optimization. Model performance was evaluated by analyzing the training loss, validation loss, and validation accuracy curves.

In binary classification tasks, the model performed exceptionally well, achieving a validation accuracy of 90.18% and an ROC-AUC value of 0.882, demonstrating strong discriminative ability between AD patients and cognitively normal individuals. However, in ternary classification tasks, the model exhibited lower validation accuracy, indicating underfitting.

The 3D VoxResNet-based MRI recognition method can provide valuable insights for early AD diagnosis and treatment.