Lung-DDPM+: Efficient Thoracic CT Image Synthesis using Diffusion Probabilistic Model

TL;DR

This work tackles the challenge of data-scarce thoracic CT analysis by introducing Lung-DDPM+, a diffusion-probabilistic model tailored for lung nodule synthesis. It couples a pulmonary DPM-solver with an effective anatomically aware sampling (EAAS) pipeline to generate nodules within $64^3$ patches conditioned on segmentation maps, dramatically improving sampling efficiency while preserving anatomical realism. The method delivers up to $8\times$ fewer TFLOPs, $21\times$ less VRAM, and $76\times$ faster sampling than previous approaches, while maintaining competitive performance in downstream lung nodule segmentation tasks and passing a Visual Turing Test by expert radiologists. By enabling high-quality, data-efficient synthetic nodules, Lung-DDPM+ has the potential to enhance training for diagnostic models and broaden the applicability of diffusion-based medical image synthesis to other tumors and modalities.

Abstract

Generative artificial intelligence (AI) has been playing an important role in various domains. Leveraging its high capability to generate high-fidelity and diverse synthetic data, generative AI is widely applied in diagnostic tasks, such as lung cancer diagnosis using computed tomography (CT). However, existing generative models for lung cancer diagnosis suffer from low efficiency and anatomical imprecision, which limit their clinical applicability. To address these drawbacks, we propose Lung-DDPM+, an improved version of our previous model, Lung-DDPM. This novel approach is a denoising diffusion probabilistic model (DDPM) guided by nodule semantic layouts and accelerated by a pulmonary DPM-solver, enabling the method to focus on lesion areas while achieving a better trade-off between sampling efficiency and quality. Evaluation results on the public LIDC-IDRI dataset suggest that the proposed method achieves 8$\times$ fewer FLOPs (floating point operations per second), 6.8$\times$ lower GPU memory consumption, and 14$\times$ faster sampling compared to Lung-DDPM. Moreover, it maintains comparable sample quality to both Lung-DDPM and other state-of-the-art (SOTA) generative models in two downstream segmentation tasks. We also conducted a Visual Turing Test by an experienced radiologist, showing the advanced quality and fidelity of synthetic samples generated by the proposed method. These experimental results demonstrate that Lung-DDPM+ can effectively generate high-quality thoracic CT images with lung nodules, highlighting its potential for broader applications, such as general tumor synthesis and lesion generation in medical imaging. The code and pretrained models are available at https://github.com/Manem-Lab/Lung-DDPM-PLUS.

Figures (5)

• Figure 1: Demonstration of low efficiency and anatomical imprecision issues in SOTA generative models for thoracic CT image synthesis.
• Figure 2: Comparasion of workflows for Lung-DDPM and Lung-DDPM+. The workflow for both models consists of four distinct steps: ① Data labeling, ② Model training, ③ Sampling from the model, and ④ Downstream task.
• Figure 3: The sampling process of the proposed method consists of three key steps: (1) Nodule semantic layout generation, (2) Effective anatomically aware sampling (EAAS) process, and (3) Fusion.
• Figure 4: Experimental results for 3D lung nodule segmentation on the NLST dataset. Subfigures (A) and (B) display the Dice and HD95 metrics, respectively. The green dashed line indicates the best performance, the red dashed line marks the worst, and the black dashed line represents the baseline metric. Note that 100%-1000% indicates the percentage of the synthetic NLST cases that are included in the training set.
• Figure 5: Demonstration of synthetic samples generated by Lung-DDPM+ and other SOTA competitors. Gray areas in the semantic layouts represent pulmonary regions, while white areas indicate lung nodule regions.