Lung-DDPM

Lung-DDPM is a 3D denoising diffusion probabilistic model that generates synthetic thoracic CT volumes conditioned on a semantic layout — a voxel map marking lung tissue and lung nodules. It was developed at the Centre de recherche du CHU de Québec and the Quebec Heart and Lung Institute, both affiliated with Université Laval, and published in IEEE Transactions on Biomedical Engineering.

The problem it addresses is specific to lung cancer screening. Deep learning for nodule detection and segmentation needs large annotated CT cohorts, but voxel-level annotation is slow and expensive, and patient privacy rules restrict sharing. Prior layout-guided generators solved this by requiring a complete semantic map of the chest, including extra-pulmonary anatomy. That requirement is impractical: in screening cohorts radiologists annotate nodules and little else, and automatic organ segmentation is unreliable across scanner types and acquisition protocols. Given only a partial layout, earlier models produce volumes with anatomically implausible mediastinum, ribcage, and body wall.

Lung-DDPM's answer is anatomically aware sampling. At every denoising step it produces the lung region from the layout-conditioned network and takes the extra-pulmonary region from a real reference scan diffused to the same noise level, then blends the two with the complement of the lung mask. The synthetic scan is therefore generated where the annotation exists and borrowed where it does not.

#Key Features

• Partial layouts are enough: Only radiologist-drawn lung and nodule masks are needed; the extra-pulmonary layout is derived as the complement of the lung mask rather than annotated.
• Anatomically aware sampling: Lung and extra-pulmonary latents are computed separately at each timestep and fused through the mask, which removes the seams and anatomical distortions that appear when a layout-guided model must invent unconditioned regions.
• Unlimited nodule-positive cases: One layout can seed arbitrarily many distinct volumes, letting a screening cohort be expanded without further annotation or patient recruitment.
• Validated on a downstream task: Image-quality metrics are reported alongside a nodule segmentation experiment, so the synthetic data is judged by whether it helps a model rather than by fidelity scores alone.

#Technical Details

The denoising network is a 3D U-Net with residual and attention blocks, adapted from the 3D-DDPM and Med-DDPM designs, taking the noised volume concatenated with the semantic layout as input. All volumes and layouts were preprocessed to 128 × 128 × 128. Training ran for 100,000 steps with an L1 objective, Adam at a learning rate of 1e-5, EMA decay 0.995, and a cosine noise schedule with 250 sampling steps, on a single A100. Data came from an internal IUCPQ cohort of 1,849 nodule-positive resection patients — 800 scans for training, 800 for synthesis — plus 366 non-contrast cases from the public LIDC-IDRI dataset as an external cohort. On the validation cohort Lung-DDPM reached an FID of 0.0047, MMD of 0.0070, and MSE of 0.0024, respectively 7.4×, 3.1×, and 29.5× better than the next-best generative baseline. A segmentation model trained on real plus synthetic scans achieved a Dice coefficient of 0.3914 and sensitivity of 0.4393, improvements of 8.8% and 18.6% over training on real scans alone.

#Applications

The direct use is data augmentation for lung cancer screening pipelines: a group with a modest annotated cohort can generate additional nodule-positive volumes and train a stronger nodule segmentation or detection model. Because synthetic scans carry no patient identity, they also offer a route to sharing training material across institutions where the underlying imaging cannot be released. The layout-conditioned interface makes controlled studies possible — nodule size, count, and location can be specified to probe how a downstream detector behaves on rare presentations. A successor, Lung-DDPM+, targets the sampling cost that limits large-scale generation.

#Impact

Lung-DDPM shows that layout-guided medical image synthesis can be made compatible with how clinical data is actually annotated, by generating only the region that is labeled and sourcing the rest from real anatomy. The absolute Dice figures reflect how difficult whole-volume nodule segmentation is rather than a ceiling on the method, and the gains come from mixing synthetic with real data, not replacing it. Training and validation draw on a single institution plus one public dataset, so generalization across scanner vendors and screening populations remains to be established; the released code is MIT-licensed, with pretrained weights distributed through a cloud drive link rather than a model hub.