Stratifying amyloid burden in early Alzheimer's disease using cascaded attention-guided vision transformer using [¹⁸F]Florbetapir PET.

Abstract

This study aims to develop a deep learning model to assist physicians in accurately classifying negative, equivocal, and positive β-amyloid (Aβ) deposition stages in Alzheimer's disease (AD). 1327 subjects from two cohorts underwent [¹⁸F]Florbetapir PET and were grouped by Aβ deposition. A cascaded attention-guided vision transformer (CA-ViT) framework was proposed to extract biologically significant regional information for fine-grained classification. To evaluate clinical utility, we assessed the diagnostic performance of physicians with and without the assistance of our proposed method. The CA-ViT model demonstrated outstanding cross-center performance, achieving accuracies of 82.8% [79.1%, 86.5%] (96% confidence interval, CI) and 78.0% [75.1%, 80.9%] in three-class classification tasks in the two cohorts, respectively. Our proposed model-assisted physicians exhibited significant improvements in diagnostic accuracy (from 56% to 66% and from 50% to 77%). The CA-ViT model effectively decodes fine-grained pathological information from [¹⁸F]Florbetapir PET imaging, enabling accurate stratification of Aβ deposition to assist physicians in early monitoring of AD. Question Deep learning has the potential to assist physicians in accurately classifying β-amyloid deposition stages in early Alzheimer's disease. Findings The proposed diagnostic model is a promising computer-aided tool for early assessment of amyloid deposition and demonstrates improved physician performance. Clinical relevance Equivocal amyloid deposition often complicates early Alzheimer's disease diagnosis and may delay optimal interventions. Our model, validated on PET scans from multiple centers, enhances the identification of these equivocal cases and improves diagnostic accuracy among less-experienced physicians.