An Appropriate Vision Transformer Model for Classifying Cognitive Impairment Progression Stages Utilizing MRI Images

Mild Cognitive Impairment (MCI) due to Alzheimer”s Disease (AD) progresses at different rates for each individual. Monitoring structural changes in the brain’s cortical areas during the early stages of cognitive impairment presents a critical opportunity for intervention before the onset of Dementia. The transition from Early Mild Cognitive Impairment (EMCI) to Late Mild Cognitive Impairment (LMCI) is a critical and irreversible change within the clinical continuum of AD, indicating a substantial worsening of cognitive decline and an elevated risk of Dementia progression. This work aims to leverage the effective feature representation of the Vision Transformer (ViT), which uses an attention mechanism, to categorize various stages of cognitive impairment. Specifically, it focuses on Early Mild Cognitive Impairment (EMCI), which is closer to Cognitive Normal (CN), and Late Mild Cognitive Impairment (LMCI), which has a higher conversion rate to Alzheimer’s Disease (AD). The work uses ViT/16 as the base model and modifies the model by implementing fine-tuned top encoder layers and classifier head, using a large volume of brain Magnetic Resonance Image (MRI) dataset, 12,000 2D images in each class, in axial orientation, helping to adapt features more effectively and enabling the model to generalize well to new unseen data and compared it by using ViT/32 as base. The study focuses on evaluating two multi-class classification tasks: distinguishing between AD, EMCI, and LMCI, and between CN, EMCI, and LMCI. Using the Alzheimer”s Disease Neuroimaging Initiative (ADNI) dataset, the proposed method achieved 89.22 ± 0.400% accuracy for the first classification task and 88.57 ± 0.298% for the second. Additionally, the model demonstrated a precision of 90.83 ± 0.670% and 90.71 ± 0.568% for the respective tasks. To visualize the model’s performance, the output probability distributions were illustrated using Kernel Density Estimation (KDE) plots and Receiver Operating Characteristic-Area Under the Curve (ROC-AUC) curves, employing both the One-vs-One (OvO) and One-vs-Rest (OvR) methods; increased KS statistics and significant p-values between classes validated the results. The results demonstrated that the modified ViT/16 model, combined with high-performance computing resources and an extensive dataset, offers a promising approach for accurately categorizing the progression of cognitive impairment stages with subtle changes. Fine-tuning certain encoder layers of a pre-trained Vision Transformer model on a large brain MRI dataset enhances performance by allowing the model to develop high-level, task-specific representations, while preserving strong low-level characteristics acquired during pretraining, making it a reliable and practical tool for assisting healthcare professionals in decision-making. Statistical significance testing shows that the proposed approach outperforms deep-learning benchmark models in classifying stages of cognitive impairment, including subtle stages within the continuum of Alzheimer’s disease, particularly between EMCI and LMCI, which is clinically significant.