Neuro-ocular amyloid characterization in Alzheimer's disease via cross-site PET-MRI and hierarchical cross-attention driven multimodal representation learning.

The eye is often described as a window to the brain and may offer a non-invasive opportunity for investigating neurodegenerative changes. Although cerebral amyloid-β (Aβ) accumulation is a well-established hallmark of Alzheimer's disease (AD), the potential of ocular regions to capture aspects of this pathology in positron emission tomography (PET) remains underexplored. Our primary objective is to conduct a radiomics-SUVR analysis to investigate the relationship between ocular regions of interest (ROI) extracted from PET imaging and regional SUVR values obtained from six brain PET regions. Our secondary objective is to address domain shifts arising from heterogeneous data distributions across imaging sites during ocular segmentation. This study included 228 participants from Chonnam National University Hospital (CNUH) and an external validation cohort of 50 participants from ADNI-4, all with paired 18F-florbetaben PET and 3D T2-weighted MRI. For the segmentation task, we propose HCA-Net, a Hierarchical Cross-Attention Network that integrates co-registered PET and MRI using a multi-scale fusion strategy, allowing the model to better handle domain shifts in cross-site imaging conditions. For the radiomics-SUVR analysis, the segmented ocular ROI was classified into Aβ⁺ and Aβ⁻ groups using a CNN model. Radiomic features were then extracted and compared with corresponding regional brain SUVRs using statistical methods. While several features showed group differences at an uncorrected threshold, only one feature remained significant after correction for multiple comparisons highlighting the need for further validation. Correlation analyses identified modest associations between ocular radiomics and regional brain SUVRs, highlighting the influence of global amyloid burden on these relationships. Overall, our study presents a multimodal cross-site segmentation framework and provides preliminary evidence of neuro-ocular associations in AD through 18F-FBB PET imaging, highlighting the need for further validation.