3D-CNN Enhanced Multiscale Progressive Vision Transformer for AD Diagnosis

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Vision Transformers (ViT) show promise in diagnosing Alzheimer's disease (AD) and mild cognitive impairment (MCI) using structural magnetic resonance images (sMRI).
• Existing ViT models face challenges with limited labeled AD-related sMRI datasets, neglecting crucial within-patch local features like brain atrophy, and exhibiting high computational complexity due to quadratic increases in patch numbers.

Purpose of the Study:

• To address the limitations of ViT in AD diagnosis, this study proposes a novel 3D-CNN Enhanced Multiscale Progressive ViT (3D-CNN-MPVT).
• The aim is to improve the accuracy and efficiency of diagnosing Alzheimer's disease and predicting mild cognitive impairment conversion by enhancing local feature learning and reducing computational overhead.

Main Methods:

• A 3D-convolutional neural network (CNN) is pre-trained on sMRI data to extract detailed local features and mitigate overfitting.
• A Multiscale Progressive ViT (MPVT) module with an integrated CNN is developed to explicitly capture within-patch interactions crucial for AD diagnosis.
• A novel stitch operation merges cross-patch features and progressively reduces the number of patches, enhancing local feature characterization while managing computational costs.

Main Results:

• The 3D-CNN-MPVT model demonstrated superior performance on large datasets (ADNI: 6610 scans, OASIS-3: 1866 scans).
• Achieved 90% accuracy in Alzheimer's disease classification and 80% accuracy in mild cognitive impairment conversion prediction with minimal preprocessing.
• Outperformed recent baseline methods in diagnostic accuracy for Alzheimer's disease and mild cognitive impairment.

Conclusions:

• The proposed 3D-CNN-MPVT effectively addresses key challenges in applying ViT to sMRI for AD and MCI diagnosis.
• The integration of 3D-CNN and MPVT with a stitch operation enhances local feature learning and reduces computational complexity.
• This approach offers a promising, highly accurate, and efficient method for neurodegenerative disease diagnosis using brain imaging data.