RLBCD: Residual-guided Latent Brownian-bridge Co-Diffusion for Anatomical-to-Metabolic Image Synthesis

Tianxu Lv, Hongnian Tian, Jiansong Fan, Yuan Liu, Lihua Li, Xiang PanIJCAI 2025 · DOI

TL;DR

This paper introduces RLBCD, a novel diffusion model for synthesizing metabolic medical images (like PET scans) from anatomical images (like CT scans). RLBCD addresses the challenge of subtle differences between image types and provides a confidence score for synthesized images, making them more reliable for clinical use.

What Problem Does This Paper Address?

Metabolic imaging, such as PET scans, is crucial for early disease diagnosis by revealing physiological changes. However, it's often expensive, slow, and involves radiation risks. Anatomical imaging, like CT scans, is more accessible but lacks this functional information. The goal of Anatomical-to-Metabolic Image Synthesis (A2MIS) is to generate metabolic images from anatomical ones, bridging this gap. Existing methods struggle with the subtle differences between anatomical and metabolic images, leading to inaccurate syntheses. Furthermore, they lack a way to quantify the confidence of the generated images, limiting their clinical applicability. This paper aims to develop a method that can accurately synthesize metabolic images while providing a reliable confidence score.

Methods

The core of RLBCD is a residual-guided latent Brownian-bridge co-diffusion process. This approach leverages two diffusion branches:

  • Residual Branch: This branch uses a Brownian-bridge diffusion process to directly learn the differences between the anatomical and metabolic domains. This is crucial for capturing subtle inter-domain variations.
  • Enhanced Branch: This branch takes the information from the residual branch and uses another Brownian-bridge diffusion process to reconstruct the target metabolic image. This branch focuses on generating high-fidelity, modality-specific details.

To effectively integrate information from these branches, RLBCD employs a Hybrid-Granularity Fusion (HGF) mechanism. This involves:

  • Semantic Alignment (SAM): This module strengthens semantic representations by aligning features from the residual branch within the encoder of the diffusion model.
  • Complementary Attention (CMA): This module integrates complementary information from the residual branch into the decoder of the diffusion model, helping to refine details.

Finally, RLBCD introduces a Latent Consistency Score (LCS). This score is derived from a pre-trained fusion network that predicts the anatomical image from the synthesized metabolic and original anatomical latent representations. The LCS acts as an indicator of the synthesized image's quality and confidence, which is crucial for clinical decision-making. The diffusion process is performed in a latent space, making it more efficient.

Datasets

The paper evaluates RLBCD on five datasets:

  • Public Datasets:
    • Head-Neck-PET-CT
    • Duke-Breast-Cancer-MRI
    • Breast-MRI-NACT-PiloT
  • In-house Datasets:
    • Chest-CTA
    • HUASHANCT-PET

These datasets cover various A2MIS tasks, including CT-to-PET, CT-to-CTA, and pre-contrast MRI-to-post-contrast MRI synthesis. The datasets are split into training, validation, and testing sets.

Key Insights and Ideas

  • Focusing on Residuals is Key: The paper highlights that directly learning the differences (residuals) between anatomical and metabolic images is more effective for A2MIS than just learning a direct mapping. This is because the subtle variations are often the most informative for metabolic function.
  • Co-Diffusion for Complementary Information: Using two diffusion branches (residual and enhanced) allows the model to capture both inter-domain differences and reconstruct target-specific details simultaneously.
  • Hybrid-Granularity Fusion for Detail Refinement: Combining semantic alignment in the encoder and complementary attention in the decoder effectively leverages features at different granularities to improve reconstruction accuracy.
  • Confidence Scores are Essential for Clinical Trust: The introduction of the Latent Consistency Score is a significant contribution, as it provides a quantifiable measure of how reliable a synthesized image is, which is paramount for medical applications.

Key Results

  • RLBCD consistently outperforms state-of-the-art baselines (including GAN-based and other diffusion models) across multiple datasets and A2MIS tasks, achieving significant gains in SSIM (up to 5.90%).
  • The ablation studies demonstrate that each component of RLBCD (Residual Branch, Enhanced Branch, HGF, and LCS) contributes to the overall performance.
  • The Latent Consistency Score shows a positive correlation with image quality, with higher LCS values corresponding to better SSIM scores.
  • In downstream tasks, synthesized images from RLBCD led to improved cancer segmentation (e.g., ~7.56% increase in DSC for breast cancer segmentation) and cardiac amyloidosis diagnosis (e.g., ~26.38% increase in AUC), validating its clinical utility.

Limitations and Future Work

  • Generalization to Diverse Data: While RLBCD shows good generalization on independent datasets, the paper notes that medical images inherently exhibit diversity due to imaging protocols and patient demographics. Further work could explore more robust domain adaptation techniques.
  • Computational Cost: Diffusion models, in general, can be computationally intensive. While RLBCD operates in a latent space, further optimizations for training and inference speed might be beneficial.
  • Clinical Validation: While the paper demonstrates improved performance in downstream diagnostic and segmentation tasks, extensive clinical validation with medical professionals would be the next logical step to fully establish its real-world impact.
  • Exploring Other Modalities: The current work focuses on specific anatomical-metabolic pairs. Future research could extend RLBCD to other medical imaging modalities and translation tasks.

Summary updated

BibTeX
@inproceedings{lv2025rlbcd,
  author = {Tianxu Lv and Hongnian Tian and Jiansong Fan and Yuan Liu and Lihua Li and Xiang Pan},
  title = {{RLBCD: Residual-guided Latent Brownian-bridge Co-Diffusion for Anatomical-to-Metabolic Image Synthesis}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/656},
}
Figure