Synthetic-to-real attentive deep learning for Alzheimer's assessment: A domain-agnostic framework for ROCF scoring.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Early Alzheimer's disease diagnosis relies on cognitive tests like the Rey-Osterrieth Complex Figure (ROCF).
• Manual ROCF scoring is time-consuming, subjective, and prone to bias.
• Limited annotated clinical data hinders deep learning model development and generalization for ROCF analysis.

Purpose of the Study:

• To develop a novel framework for automated ROCF scoring using synthetic data.
• To address challenges of limited data and domain shift in deep learning models for cognitive assessments.
• To improve the accuracy, efficiency, and fairness of ROCF scoring for Alzheimer's disease diagnosis.

Main Methods:

• A lightweight data synthesis pipeline generating diverse, annotated ROCF drawings.
• ROCF-Net, a deep learning model designed for cross-domain ROCF scoring, addressing texture and line artifact variations.
• A novel line-specific attention mechanism within ROCF-Net to enhance scoring accuracy.

Main Results:

• The framework generates realistic ROCF drawings reflecting Alzheimer's abnormalities efficiently.
• ROCF-Net achieved state-of-the-art performance (MAE 3.53, PCC 0.86) across diverse datasets.
• Models trained on synthetic data showed comparable generalization to those trained on real clinical data, with minimal performance differences.

Conclusions:

• Introduced a cost-effective synthetic ROCF data generation pipeline.
• Developed a domain-agnostic model for automated ROCF scoring.
• Integrated a lightweight attention mechanism for transparent clinical scoring alignment.
• Proposed a bias-aware framework using synthetic data to mitigate demographic disparities in cognitive assessments.