Multi-slice MRI-based Alzheimer’s disease classification using EfficientNet and transformer architecture

Alzheimer’s disease (AD) is a degenerative brain disorder that results in cognitive decline, memory deterioration, and dementia. It is crucial to diagnose the condition in an early and accurate manner for optimal management of patients. MRI has been identified as a helpful tool in recognizing structural brain alterations linked with AD; nevertheless, automatic recognition of spatial relations between MRI slices poses a great difficulty. This paper aims at developing a new deep learning algorithm based on multi-slice MRI for the classification of CN, MCI, and AD conditions. Our architecture incorporates EfficientNet-B0 and transformer encoder to extract spatial features from MRI slices and capture inter-slice dependencies and relationships, respectively. Consecutive MRI slices obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database were grouped into multi-slice inputs. Cross-entropy loss function and AdamW optimization approach were used during the training process. Model performance was assessed via accuracy, precision, recall, F1-score, confusion matrices, and ROC curve plots. The proposed model attained an overall accuracy of 88 and an average Area Under the Curve (AUC) value of 0.946. Class-wise performance of the model showed good performance for all categories, where the precision values were 0.898, 0.833, and 0.836 for CN, MCI, and AD, respectively. The model was able to learn structural patterns related to disease progression while having good performance on various diagnostic classes. The combination of EfficientNet-B0 and Transformer attention techniques provides an efficient way of learning features from within slice and between slice MRI scans for Alzheimer’s disease classification. The multi-slice model shows promising results in classifying CN, MCI, and AD patients.