Hybrid Vision Transformer-CNN Framework for Alzheimer's Disease Cell Type Classification: A Comparative Study with...

Area of Science:

• Computational biology
• Neuroscience
• Artificial intelligence

Background:

• Accurate identification of Alzheimer's disease (AD) cellular characteristics from microscopy is crucial for understanding neurodegeneration.
• Computational approaches often overlook microscopy-based cell type classification, focusing instead on macroscopic neuroimaging.

Purpose of the Study:

• To develop and evaluate a hybrid vision transformer-convolutional neural network (ViT-CNN) framework for classifying AD-related cell types from phase-contrast microscopy images.
• To compare the performance of the proposed ViT-CNN model against conventional CNN architectures and large language models (LLMs) in a cell classification task.

Main Methods:

• A hybrid ViT-CNN framework integrating DeiT-Small and EfficientNet-B7 was developed.
• The model was trained to classify three AD-related cell types: astrocytes, cortical neurons, and SH-SY5Y neuroblastoma cells.
• Comparative evaluation included standalone CNNs (DenseNet, ResNet, InceptionNet, MobileNet) and prompt-based LLMs (GPT-5, GPT-4o, Gemini 2.5-Flash) using zero-shot, few-shot, and chain-of-thought prompting.

Main Results:

• The proposed hybrid ViT-CNN model achieved a test accuracy of 61.03% and a macro F1 score of 61.85%.
• The hybrid model outperformed standalone CNN baselines and prompt-only LLM approaches, particularly under data-limited conditions.
• Results indicate that combining convolutional inductive biases with transformer-based global context modeling enhances generalization for cellular microscopy classification.

Conclusions:

• The hybrid ViT-CNN framework demonstrates improved generalization for classifying AD-related cell types from microscopy images.
• This study serves as a proof of concept, highlighting the potential of integrating transformers and CNNs for cellular analysis.
• Future research directions include domain-specific pretraining, multimodal data integration, and explainable AI for AD research.