Abstract

Three-stage process for image synthesis and disease classification: Stage 1: MRI preprocessing for standardization. Stage 2: Aβ-PET synthesis from MRI using the IL-CLDM method. Stage 3: Multimodal learning for disease classification using MRI and synthesized Aβ-PET.

β-amyloid (Aβ) deposition, typically observed through Aβ-PET, is a key biomarker for early-onset dementia. However, Aβ-PET is costly and involves radiation, limiting its use in large-scale screening compared to MRI. Since Aβ-PET scans only indicate Aβ deposition status, it is highly valuable to predict amyloid status using MRI to reduce the need for Aβ-PET. To this end, we propose an image-and-label conditional latent diffusion model (IL-CLDM) to synthesize Aβ-PET scans from MRI, enhancing key shared information for MRI-based Aβ status classification. IL-CLDM features two conditioning modules: 1) an image conditioning module, to extract meaningful features from source MRI scans to provide structural information, and 2) a label conditioning module, to guide the alignment of generated scans to the diagnosed label. Experiments on a clinical dataset show that IL-CLDM outperforms five widely used models, and the synthesized Aβ-PET scans significantly aid Aβ classification.

Methodology

Illustration of our proposed IL-CLDM. It first resorts to an adversarial autoencoder to compress each Aβ-PET scan into a lower-dimensional latent representation, and then maps the MRI scan to the target Aβ-PET scan on the latent space.

We propose an image-and-label conditional latent diffusion model (IL-CLDM) to synthesize disease-sensitive Aβ-PET scans from corresponding MRI scans, thereby enhancing the accuracy of MRI-based Aβ status classification. IL-CLDM first resorts to an adversarial autoencoder to compress Aβ-PET scans into lower-dimensional latent representations, preserving critical medical information. The model then learns the relationship between Aβ-PET and MRI in latent space, rather than the original image space. By incorporating image and label conditioning modules, IL-CLDM enables joint learning of image synthesis and disease diagnosis, ensuring the synthetic Aβ-PET scans are both realistic and diagnostically relevant.

Results

Visualization of real and synthetic Aβ-PET scans generated by different methods. The top three rows correspond to positive amyloid status, while the bottom three correspond to negative amyloid status.

Visualization of real and synthetic Aβ-PET scans generated by two different variants of our model (with w and without w/o label conditioning module). The respective error maps are also provided.

Performances of different methods on the task of Aβ status classification.

Performances of different methods on the task of AD status classification.

Citation

@ARTICLE{10752348, 
  title={Image-and-Label Conditioning Latent Diffusion Model: Synthesizing Aβ-PET from MRI for Detecting Amyloid Status},
  author={Ou, Zaixin and Pan, Yongsheng and Xie, Fang and Guo, Qihao and Shen, Dinggang},  
  journal={IEEE Journal of Biomedical and Health Informatics},  
  pages={1--11},  
  year={2024}
 }