(주) 뷰노

(주) 뷰노

문의사항 남기기
KR EN
VUNO MED INC.
Publications
BACK TO LIST
AAIC 2025

Predicting Amyloid Burden Using a Masked Multimodal-Multitask Deep Learning Framework with Latent Diffusion-based Synthetic PET

  • Jul. 2025

Background: Alzheimer’s disease (AD) is characterized by pathological amyloid-β accumulation. Positron emission tomography (PET) imaging remains a widely accepted gold standard for quantifying amyloid burden via standardized uptake value ratios (SUVRs); however, its prohibitive cost and limited availability restrict broader clinical adoption. Although recent deep learning models show promise in predicting amyloid burden, they often require complete clinical data or actual PET images—conditions rarely satisfied in routine clinical practice. To overcome these barriers, we propose a masked multimodal–multitask deep learning framework that integrates synthetic PET scans to improve amyloid burden prediction under real-world data constraints.

 

Method: We analyzed 2,043 longitudinal observations from 968 ADNI-2 and ADNI-3 participants (mean age: 72 years; male/female: 491/477). Each observation included T1-weighted, T2*-weighted, and FLAIR MRI, AV45-PET, demographic data, APOE4 status, cognitive assessments, and CSF biomarkers. Our approach involved two main steps:

1.     Synthetic PET generation: A latent diffusion model (LDM) was trained to generate synthetic AV45-PET scans from MRI sequences.

2.     Amyloid burden prediction: A deep learning network integrated these synthetic PET images with available clinical data through a masked embedding attention mechanism to explicitly handle missing inputs and predict both continuous amyloid SUVRs and masked clinical features.

We compared our method against three baselines: (i) MRI-only, (ii) non-imaging–only, and (iii) a multimodal model without synthetic PET. Performance on a held-out test set was evaluated using mean absolute error (MAE) for SUVR prediction and area under the curve (AUC) for amyloid positivity (SUVR > 1.11).

 

Result: For SUVR prediction, the baseline models achieved MAEs of 0.20, 0.13, and 0.13, whereas our approach reached 0.11. For amyloid positivity classification, the baselines yielded AUCs of 0.48, 0.90, and 0.91, while our model attained 0.93.

 

Conclusion: By reducing reliance on cost-intensive PET scans and handling missing data more effectively, our framework could substantially broaden access to early AD screening in diverse clinical settings. Future work will involve testing in larger, more varied cohorts, applying different tracers, and further evaluate its real-world impact on AD management.
 
 

Tag in

#medical_image

#VUNO Med®-DeepBrain®

이전글
Revolutionizing Radiology Reporting With Artificial Intelligence-Based Voice Recognition: A Pilot Study on Lumbar Spine Magnetic Resonance Imaging 2025. 08. 04
다음글
Discriminative Feature Learning for Lacune Detection in 2D T2-FLAIR Images using Supervised Contrastive Learning 2025. 08. 04
List

이 웹사이트는 원활한 서비스 제공을 위하여 당사 개인정보 처리방침에 명시된 쿠키를 사용합니다.

만일 당사의 쿠키 사용 정책에 대해 더 알고 싶으신 경우에는 개인정보 처리방침 을 참고하여 주시기 바랍니다.