1373-UR Latent Cross-Modal Completion for Alzheimer’s Disease Diagnosis with Incomplete Multimodal Data
Location
https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php
Document Type
Event
Start Date
20-11-2026 2:00 PM
Description
Multimodal data provide complementary information for Alzheimer’s disease (AD) diagnosis, but acquiring all modalities can be impractical due to cost, accessibility, and patient burden. Real-world datasets also frequently contain incomplete observations, limiting the use of multimodal models requiring all modalities. To address these challenges, we propose framework that learns shared latent representations from MRI-derived features, amyloid PET, and clinical information. Rather than reconstructing missing modalities in their original feature spaces, the model learns cross-modal mappings in a shared latent space, enabling available modalities to estimate representations of missing ones. Contrastive learning preserves shared cross-modal information, while feature- and modality-level attention mechanisms adaptively integrate complementary information for diagnosis. We evaluate the model on three independent datasets (ADNI, AIBL, and OASIS) across three binary and one three-class classification tasks, with simulated missing-modality rates of 0% to 50%. Results show that the proposed model maintains diagnostic performance under varying levels of modality incompleteness, although robustness differs across datasets and tasks. These findings demonstrate the potential of latent cross-modal completion for AD diagnosis when complete multimodal data are unavailable.
1373-UR Latent Cross-Modal Completion for Alzheimer’s Disease Diagnosis with Incomplete Multimodal Data
https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php
Multimodal data provide complementary information for Alzheimer’s disease (AD) diagnosis, but acquiring all modalities can be impractical due to cost, accessibility, and patient burden. Real-world datasets also frequently contain incomplete observations, limiting the use of multimodal models requiring all modalities. To address these challenges, we propose framework that learns shared latent representations from MRI-derived features, amyloid PET, and clinical information. Rather than reconstructing missing modalities in their original feature spaces, the model learns cross-modal mappings in a shared latent space, enabling available modalities to estimate representations of missing ones. Contrastive learning preserves shared cross-modal information, while feature- and modality-level attention mechanisms adaptively integrate complementary information for diagnosis. We evaluate the model on three independent datasets (ADNI, AIBL, and OASIS) across three binary and one three-class classification tasks, with simulated missing-modality rates of 0% to 50%. Results show that the proposed model maintains diagnostic performance under varying levels of modality incompleteness, although robustness differs across datasets and tasks. These findings demonstrate the potential of latent cross-modal completion for AD diagnosis when complete multimodal data are unavailable.