Multi-Conditioned Denoising Diffusion Probabilistic Model (mDDPM) for Medical Image Synthesis

Arjun Krishna, Ge Wang, Klaus MuellerarXiv 2024 · DOI · arXiv

Abstract

High-fidelity synthetic medical images are essential for training robust deep learning models, yet existing generative approaches often struggle with anatomical accuracy and precise control. This paper introduces the Multi-Conditioned Denoising Diffusion Probabilistic Model (mDDPM), a framework that leverages multi-factor guidance during the diffusion sampling process. By incorporating simultaneous constraints such as anatomy maps and reference slices, mDDPM generates high-resolution lung CT images that maintain strict anatomical integrity. Experimental results demonstrate that mDDPM outperforms state-of-the-art GANs and unguided diffusion models in realism and structural similarity. Notably, the generated images successfully passed a "Visual Turing Test," where expert radiologists were unable to distinguish synthetic scans from real ones.

Introduction

Deep learning in medical imaging is frequently hampered by the scarcity of high-quality, annotated datasets. Privacy regulations and the high cost of manual labeling by experts create a significant bottleneck. While synthetic data generation via Generative Adversarial Networks (GANs) has been explored, these models often produce "hallucinated" structures that lack clinical validity. Furthermore, unconditional models offer little control over specific anatomical features or pathological presentations.

To address these challenges, we present mDDPM, which extends the capabilities of Denoising Diffusion Probabilistic Models (DDPMs). By guiding the iterative denoising process with multiple simultaneous conditions, the model achieves a high degree of control and anatomical realism, effectively bridging the gap between synthetic and real-world clinical data.

Methodology

The mDDPM framework builds upon the iterative noise-reversal process of standard DDPMs. The core innovation lies in a novel sampling strategy that integrates multiple guidance signals without requiring extensive retraining for each new condition.

Multi-Condition Guidance and the "Super-Subset" Effect

Instead of training a conditional model from scratch, mDDPM utilizes a pre-trained unconditional model. During sampling, the process is "steered" using multiple reference images, such as segmentation maps or specific CT slices. This approach allows the model to sample from the intersection of multiple distributions. As illustrated in Figure 1, applying multiple conditions (e.g., Condition A + Condition B) effectively narrows the sampling space to a "super-subset" where all constraints are satisfied, leading to superior anatomical consistency.

Conceptual diagram of multi-conditioned sampling showing the intersection of different condition distributions.

Iterative Latent Variable Refinement

The guidance is implemented using linear low-pass filters ($\phi_n$) applied to latent variables at each denoising step. This technique, inspired by Iterative Latent Variable Refinement (ILVR), ensures that the generated image retains the coarse structural features of the guidance images while the diffusion model synthesizes fine-grained textures and medically plausible details.

Full Hounsfield Unit Range

A critical technical detail is the model's training on the full Hounsfield Unit (HU) range (-1000 to 1000 HU). Unlike standard image generation models that normalize pixel values to a narrow 0-255 range, mDDPM preserves the full dynamic range of CT data. This allows clinicians to view the generated images in various clinical windows—such as bone, lung, or soft tissue—without loss of diagnostic detail.

Experimental Results

The model was trained on a dataset of 5,000 low-dose 2D lung CT scans and evaluated against 10,000 generated images. Performance was compared to state-of-the-art architectures including StyleGAN2 and the Diffusion Transformer (DiT).

Quantitative Performance

mDDPM achieved a Fréchet Inception Distance (FID) of 69.85, outperforming StyleGAN2 (72.31) and unguided diffusion (83.24). Additionally, it reached a Structural Similarity Index (SSIM) of 0.45, the highest among all tested models, indicating superior capture of structural nuances.

The Visual Turing Test

To assess clinical realism, three expert radiologists performed a blinded evaluation of 30 images (15 real, 15 synthetic). The radiologists were asked to classify each image as "Real" or "Fake." The results, shown in Figure 4, reveal an overall accuracy of 45.56%. Since this is lower than the 50% expected by random guessing, it demonstrates that the synthetic images are indistinguishable from real scans even to trained eyes.

Confusion matrices for three radiologists showing their inability to reliably distinguish real from fake images.

Discussion and Future Work

The success of mDDPM highlights that anatomical accuracy is the primary factor in the perceived realism of medical images. By enforcing these constraints through multi-conditioning, the model achieves a level of fidelity suitable for clinical applications.

However, limitations remain. The current strategy is optimized for up to three or four simultaneous conditions; exceeding this may lead to an empty intersection of distributions. Future research will focus on extending the framework to 3D volumes to ensure volumetric consistency and applying the guidance mechanism to synthesize rare pathologies, which would provide invaluable training data for diagnostic AI systems.

Conclusion

The mDDPM framework represents a significant step forward in medical image synthesis. By combining the generative power of diffusion models with flexible, multi-factor guidance, it produces anatomically accurate, high-resolution lung CT images that pass the ultimate test of clinical realism.

Summary updated

BibTeX
@misc{krishna2024multi,
  author = {Arjun Krishna and Ge Wang and Klaus Mueller},
  title = {{Multi-Conditioned Denoising Diffusion Probabilistic Model (mDDPM) for Medical Image Synthesis}},
  year = {2024},
  eprint = {2409.04670},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2409.04670},
}
Figure