https://github.com/DataScience-Lab-Yonsei/25-1_DSL_Modeling_CV_PET_Time_Reduction

Project summary

<aside>

This project, titled "SCOPE: Slice-Consistent PET Reconstruction with 2D BBDM," aims to reconstruct high-quality, full-duration Amyloid PET images from significantly shortened scan times.

• Objective: To reduce the standard PET scan acquisition time from 20 minutes to 4 minutes without compromising diagnostic accuracy.
• Methodology: The team utilized a Brownian Bridge Diffusion Model (BBDM) adapted for slice-consistent reconstruction. To address the spatial inconsistencies often found in 2D generative models, they integrated two key techniques:
  • Style Key Conditioning (SKC): To guide denoising and ensure stylistic uniformity.
  • Inter-Slice Trajectory Alignment (ISTA): To encourage trajectory smoothness and anatomical coherence across adjacent slices.
• Results: The BBDM model equipped with SKC and ISTA outperformed GAN-based baselines (CycleGAN, CUT, DCLGAN) and standard diffusion models. It achieved the highest scores in Image Quality Metrics (PSNR, SSIM) and produced reconstructions visually and diagnostically comparable to full-time scans, making it viable for Alzheimer's diagnosis. </aside>

Project background

<aside>

The project addresses significant clinical limitations in the current workflow for diagnosing Alzheimer's disease via neuroimaging.

• Clinical Context: Amyloid PET imaging using radioactive tracers like 18F-florbetaben (FBB) is essential for visualizing amyloid-beta plaques in the brain, which appear as bright regions in the outer cortex of Alzheimer's patients.
• The Problem: Standard Amyloid PET scans require long acquisition times, typically up to 20 minutes. This duration causes several issues:
  • Patient Discomfort: It is difficult for patients, particularly the elderly, to remain still.
  • Artifacts: Patient movement leads to motion-induced artifacts that degrade image quality.
  • Resource Constraints: Long scans increase costs and limit scanner throughput.
• Technical Challenges: While Deep Learning (GANs and diffusion models) can reconstruct images from short scans, existing approaches face a trade-off.
  • 3D Models: Capture spatial continuity well but are computationally heavy and resource-intensive.
  • 2D Models: Are efficient but often produce "slice inconsistency," where stacked slices look misaligned or incoherent (as seen in Figure 1 of the report).
• Motivation: The team sought to combine the computational efficiency of 2D architectures with the spatial coherence of 3D models by adapting a slice-consistent framework originally designed for CT-to-MRI translation </aside>

Parts by me

<aside>

For this project, I implemented a lot more models compared to the models inside the report:

DDPM, DDIM,

DDBM: Denoising Diffusion Bridge Models

and finally

BBDM: Image-to-image Translation with Brownian Bridge Diffusion Models

compared a bunch of metrics and results, and had tried to integrate the model.

</aside>