Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Lesion-Aware Post-Training of Latent Diffusion Models for Synthesizing Diffusion MRI from CT Perfusion

Junhyeok Lee, Hyunwoong Kim, Hyungjin Chung, Heeseong Eom, Joon Jang, Chul-Ho Sohn, Kyu Sung Choi

TL;DR

This work tackles the challenge of synthesizing diffusion MRI from CT perfusion in acute ischemic stroke by introducing a lesion-aware post-training framework for latent diffusion models. Building on a base conditional LDM with a VQGAN encoder/decoder, the method adds image-space losses that target both global image fidelity and ischemic lesion regions, resulting in improved lesion delineation and overall image quality. On an 817-patient dataset, the proposed cLDM-PT consistently outperforms GAN-based and other latent diffusion baselines across MAE, PSNR, MS-SSIM, and FID, while achieving the best lesion MAE. The framework’s demonstrated gains and its successful generalization to another latent diffusion variant suggest broad applicability to medical image translation tasks beyond diffusion MRI synthesis from CT perfusion.

Abstract

Image-to-Image translation models can help mitigate various challenges inherent to medical image acquisition. Latent diffusion models (LDMs) leverage efficient learning in compressed latent space and constitute the core of state-of-the-art generative image models. However, this efficiency comes with a trade-off, potentially compromising crucial pixel-level detail essential for high-fidelity medical images. This limitation becomes particularly critical when generating clinically significant structures, such as lesions, which often occupy only a small portion of the image. Failure to accurately reconstruct these regions can severely impact diagnostic reliability and clinical decision-making. To overcome this limitation, we propose a novel post-training framework for LDMs in medical image-to-image translation by incorporating lesion-aware medical pixel space objectives. This approach is essential, as it not only enhances overall image quality but also improves the precision of lesion delineation. We evaluate our framework on brain CT-to-MRI translation in acute ischemic stroke patients, where early and accurate diagnosis is critical for optimal treatment selection and improved patient outcomes. While diffusion MRI is the gold standard for stroke diagnosis, its clinical utility is often constrained by high costs and low accessibility. Using a dataset of 817 patients, we demonstrate that our framework improves overall image quality and enhances lesion delineation when synthesizing DWI and ADC images from CT perfusion scans, outperforming existing image-to-image translation models. Furthermore, our post-training strategy is easily adaptable to pre-trained LDMs and exhibits substantial potential for broader applications across diverse medical image translation tasks.

Lesion-Aware Post-Training of Latent Diffusion Models for Synthesizing Diffusion MRI from CT Perfusion

TL;DR

This work tackles the challenge of synthesizing diffusion MRI from CT perfusion in acute ischemic stroke by introducing a lesion-aware post-training framework for latent diffusion models. Building on a base conditional LDM with a VQGAN encoder/decoder, the method adds image-space losses that target both global image fidelity and ischemic lesion regions, resulting in improved lesion delineation and overall image quality. On an 817-patient dataset, the proposed cLDM-PT consistently outperforms GAN-based and other latent diffusion baselines across MAE, PSNR, MS-SSIM, and FID, while achieving the best lesion MAE. The framework’s demonstrated gains and its successful generalization to another latent diffusion variant suggest broad applicability to medical image translation tasks beyond diffusion MRI synthesis from CT perfusion.

Abstract

Image-to-Image translation models can help mitigate various challenges inherent to medical image acquisition. Latent diffusion models (LDMs) leverage efficient learning in compressed latent space and constitute the core of state-of-the-art generative image models. However, this efficiency comes with a trade-off, potentially compromising crucial pixel-level detail essential for high-fidelity medical images. This limitation becomes particularly critical when generating clinically significant structures, such as lesions, which often occupy only a small portion of the image. Failure to accurately reconstruct these regions can severely impact diagnostic reliability and clinical decision-making. To overcome this limitation, we propose a novel post-training framework for LDMs in medical image-to-image translation by incorporating lesion-aware medical pixel space objectives. This approach is essential, as it not only enhances overall image quality but also improves the precision of lesion delineation. We evaluate our framework on brain CT-to-MRI translation in acute ischemic stroke patients, where early and accurate diagnosis is critical for optimal treatment selection and improved patient outcomes. While diffusion MRI is the gold standard for stroke diagnosis, its clinical utility is often constrained by high costs and low accessibility. Using a dataset of 817 patients, we demonstrate that our framework improves overall image quality and enhances lesion delineation when synthesizing DWI and ADC images from CT perfusion scans, outperforming existing image-to-image translation models. Furthermore, our post-training strategy is easily adaptable to pre-trained LDMs and exhibits substantial potential for broader applications across diverse medical image translation tasks.

Paper Structure

This paper contains 18 sections, 8 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Base Conditional Latent Diffusion Model
  4. Lesion-Aware Image Space Objectives for LDM Post-Training
  5. Experiments
  6. Experimental Setup
  7. Dataset
  8. Implementation Details
  9. Evaluation Metrics
  10. Baselines
  11. Experimental Results
  12. Quantitative Results
  13. Qualitative Evaluation
  14. Impact of Loss Weights
  15. Generalization to Other Latent Diffusion Models
  16. ...and 3 more sections

Figures (3)

  • Figure 1: Previous models fail to accurately depict ischemic stroke lesion in diffusion MRI synthesized from CT perfusion. Our model (right) shows higher lesion conspicuity (red) and enhanced image fidelity, highlighted with grey-white matter differentiation (blue).
  • Figure 2: Overview of our post-training framework. During post-training, in addition to the latent objective for the conditional LDM $\mathcal{L}_{latent}$, we introduce medical image space objectives $\mathcal{L}_{image}$, $\mathcal{L}_{lesion}$ to enhance overall image quality and lesion conspicuity.
  • Figure 3: Visualization of synthesized diffusion MRI images from CTP images in acute ischemic stroke patients. Our model with post-training (cLDM-PT) excels in lesion delineation (red arrows), accurately depicting ischemic stroke lesions with restricted diffusion (red contour) based on hypo-perfused regions in source CTP images. (Top) A case with infarct core in the left inferior frontal area. (Middle) A case of acute ischemic stroke by large vessel occlusion in the right middle cerebral artery. (Bottom) A case of acute infarction in the left occipital lobe.