Skip to content

Latest commit

 

History

History
77 lines (65 loc) · 3.47 KB

README.md

File metadata and controls

77 lines (65 loc) · 3.47 KB

Latent Diffusion Model for PET Resolution Recovery (LDM-RR)

Enhancing Amyloid PET Quantification: MRI-guided Super-Resolution using Latent Diffusion Models
Life 2024, Special Issue: Alzheimer’s Disease: Recent Developments in Pathogenesis, Diagnosis, and Therapy

Overview

A novel approach to improve Amyloid quantification and address partial volume effect (PVE) in PET using MRI-guided super-resolution with latent diffusion model. The diffusion model is trained on synthetic data simulated from real MRI scans using PET imaging physics.

Qualitative Results

This repository contains:

  • An inference script ./generate.py to synthesize high-resolution FBP scans given an input low-resolution FBP scan and matching T1w MRI scan.
  • Trained model checkpoints in ./checkpoints and a sample FBP and T1w MRI scan ./samples directory.
  • Training scripts in ./src to train modality-specific AutoEncoderKL (compression) models and a latent diffusion model (LDM) for super-resolution.

Installation

Instructions to install MONAI can be found here and the appropriate version of Pytorch using locally. Packages used in the current version of this code.

Dataset

You should structure your dataset in the following way, where images contain low-resolution images, mrs contain matching MRI scans and targets contain ground-truth high-resolution scans (digital phantoms here). For an example checkout ./src/spdp_fbp folder. All modality images are 256x256x256 in dimensions (sample images for reference in ./src/sample directory).

dataset/
  ├── train
    ├──images
      ├──xxx.nii.gz
      ├──...
  ├── val
    ├──mrs
      ├──yyy.nii.gz
      ├──...
  ├── test
    ├──targets
      ├──zzz.nii.gz
      ├──...

Inference

Training and inference scripts are in ./src folder. Sample inputs ./sample and generated outputs ./results are also provided.

python3 generate.py --input_folder ./sample --output_folder ./results

Training

To train the compression models:

python3 train_autoencoderKL.py --dataset /path/to/data --modality sp --epochs 60 --batch_size 2

Use modality = [sp, dp, or mr] to train models for simulated PET (low-res), digital phantom (high-res), and MRI (high-res), respectively.

For training the latent diffusion model:

python3 train_ldm_fbp_sp2dp.py --dataset /path/to/data --modality sp --epochs 60 --batch_size 2

Citation

If you find anything in the paper or repository useful, please consider citing:

@article{shah2024enhancing,
  title={Enhancing Amyloid PET Quantification: MRI-Guided Super-Resolution Using Latent Diffusion Models},
  author={Shah, Jay and Che, Yiming and Sohankar, Javad and Luo, Ji and Li, Baoxin and Su, Yi and Wu, Teresa and Alzheimer’s Disease Neuroimaging Initiative},
  journal={Life},
  volume={14},
  number={12},
  pages={1580},
  year={2024},
  publisher={MDPI}
}

Acknowledgments

This work is supported by the National Institute on Aging (NIA) grant RF1AG073424, P30AG072980, the Arizona Department of Health Services (ADHS) and the state of Arizona (ADHS Grant No. CTR057001).

Feel free to reach out in case of any queries jayshah.me