LAND: Lung and Nodule Diffusion for 3D Chest CT Synthesis with Anatomical Guidance
Anna Oliveras Roger Marí Rafael Redondo Oriol Guardià, Ana Tost Bhalaji Nagarajan Carolina Migliorelli Vicent Ribas Petia Radeva · 2025
Read LAND: Lung and Nodule Diffusion for 3D Chest CT Synthesis with Anatomical Guidance by Anna Oliveras Roger Marí Rafael Redondo Oriol Guardià, Ana Tost Bhalaji Nagarajan Carolina Migliorelli Vicent Ribas Petia Radeva (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-18446.txt.
Introduction
Anna Oliveras and her eight co-authors introduce LAND, a model that creates artificial three-dimensional chest CT scans for training and teaching. Real scans are costly to share, so synthetic scans guided by anatomy maps could help. LAND is a diffusion model, meaning it builds images by removing noise step by step, and it takes three-dimensional masks that outline the lungs and any nodules as its guide.
The model produces full chest volumes at one millimeter detail on a single mid-range graphics card, far cheaper than earlier approaches. Experiments carried a warning: guiding the model with nodule masks alone produced anatomically wrong scans, while adding the whole lung shape fixed the problem. With both guides, LAND can make varied scans with or without nodules of different traits, giving trainers and students fresh material without exposing patient data.
Catalog record
| CREATOR | Anna Oliveras Roger Marí Rafael Redondo Oriol Guardià, Ana Tost Bhalaji Nagarajan Carolina Migliorelli Vicent Ribas Petia Radeva |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 10 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-18446
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
LAND: Lung and Nodule Diffusion for 3D Chest CT Synthesis with Anatomical Guidance Anna Oliveras Roger Marí Rafael Redondo Oriol Guardià, Ana Tost Bhalaji Nagarajan Carolina Migliorelli Vicent Ribas Petia Radeva Abstract This work introduces a new latent diffusion model to generate high-quality 3D chest CT scans conditioned on 3D anatomical masks. The method synthesizes volumetric images of size 256x256x256 at 1 mm isotropic resolution using a single mid-range GPU, significantly lowering the computational cost compared to existing approaches. The conditioning masks delineate lung and nodule regions, enabling precise control over the output anatomical features. Experimental results demonstrate that conditioning solely on nodule masks leads to anatomically incorrect outputs, highlighting the importance of incorporating global lung structure for accurate conditional synthesis. The proposed approach supports the generation of diverse CT volumes with and without lung nodules of varying attributes, providing a valuable tool for training AI models or healthcare professionals. 1 Introduction Deep learning in medical imaging is hindered by the scarcity of large, diverse datasets, constrained by privacy concerns, costs, and the need for expert labeling. Synthetic data offers a promising solution, with potential impact in critical areas such as lung cancer, the leading cause of cancer-related deaths [ 3 ] . Diffusion models [ 11 ] have emerged as the most powerful generative framework, surpassing VAEs [ 16 ] and GANs [ 7 ] in realism and stability [ 5 , 19 ] . However, scaling diffusion models to large synthetic volumes such as CT scans remains challenging due to extreme computational demands [ 14 ] . Recent methods have explored efficiency trade-offs. Previous Latent Diffusion Models (LDMs) [ 22 ] for 3D synthesis use autoencoders for data compression, but are often limited in resolution [ 21 , 15 ] . PatchDDM [ 2 ] and WDM [ 6 ] bypass autoencoders with subvolume or wavelet representations but still require large GPU memory. NVIDIA’s LDM MAISI [ 8 ] attains the highest resolution to date ( 512\times 512\times 768 ), but demands 49.7GB GPU memory, unaffordable for most users. We introduce LAND (Lung-And-Nodule-Diffusion), a memory-efficient latent diffusion model for 3D chest CT synthesis. It generates 256^{3} volumes at 1mm resolution on a single 20GB GPU, uses lung and nodule masks for anatomical conditioning, and controls nodule texture for realistic pathological diversity. LAND combines computational efficiency with fine-grained anatomical control to achieve state-of-the-art (SOTA) high-resolution volume synthesis with practical hardware requirements. Figure 1: LAND uses a 3D VAE to encode CT volumes into a latent space, where a 3D U-Net performs diffusion on the latent samples \mathbf{x} , optionally conditioned on anatomical masks \mathbf{m} (lungs/nodules). 2 Method LAND is a latent diffusion model comprising a 3D U-Net and a 3D VAE architecture (Fig. 1 ). 3D VAE A 3D VAE encodes input CT images into latent representations compressing 4\times the spatial resolution and expanding 4\times the feature dimensionality: each 256\times 256\times 256 volume is encoded as a 64\times 64\times 64\times 4 latent sample. We adopt a lightweight variant of the MAISI architecture [ 8 ] , using 3 resolution levels with one residual block per level and the same number of channels in both encoder and decoder. The VAE is trained using a combination of an L_{1} loss \mathcal{L}_{\text{MAE}} , a perceptual similarity loss \mathcal{L}_{\text{LPIPS}} , an adversarial loss \mathcal{L}_{\text{ADV}} and a Kullback-Leibler term \mathcal{L}_{\text{KL}} : \mathcal{L}_{\text{VAE}}=\mathcal{L}_{\text{MAE}}(\mathbf{x},\hat{\mathbf{x}})+\mathcal{L}_{\text{LPIPS}}(\mathbf{x},\hat{\mathbf{x}})+\mathcal{L}_{\text{ADV}}(\mathbf{x},\hat{\mathbf{x}})+\mathcal{L}_{\text{KL}}(\mathcal{E}(\mathbf{x})) , where \hat{\mathbf{x}}=\mathcal{D}(\mathcal{E}(\mathbf{x})) is the reconstructed encoded-decoded volume. \mathcal{L}_{\text{MAE}} and \mathcal{L}_{\text{LPIPS}} enforce numerical and perceptual fidelity [ 25 ] , while \mathcal{L}_{\text{KL}} regularizes the latent space [ 16 ] and \mathcal{L}_{\text{ADV}} prevents unrealistic artifacts [ 8 , 7 ] . 3D U-Net The denoising network is a 3D U-Net with 5 resolution levels and two residual blocks per level. Additive skip connections [ 2 ] reduce memory load while preserving spatial information. To enhance conditional generation, cross-attention modules [ 22 , 17 ] re-inject conditioning masks (if any) at multiple resolution levels. Training uses velocity prediction [ 23 ] , enabling the U-Net to learn denoising by estimating a linear combination of clean latent and added noise, which stabilizes training and improves high-resolution synthesis [ 12 , 23 ] . A linear noise schedule is applied, and training follows a Min-SNR- \gamma loss weighting [ 9 ] to balance timestep contributions by signal-to-noise ratio: \mathcal{L}_{\text{min-SNR}}=\gamma(\text{SNR}_{t})|\hat{\mathbf{v}}_{t}(\mathbf{z}_{t},\mathbf{m})-\mathbf{v}_{t}|^{2}, where \mathbf{z}_{t} is the noisy latent, \mathbf{m} the conditioning mask, \gamma(\cdot) the Min-SNR weight, and \mathbf{v}_{t},\hat{\mathbf{v}}_{t} the target and predicted velocities. To ensure anatomical plausibility in 3D, LAND can be conditioned on masks \mathbf{m} covering lungs and nodules. Unlike prior 2D work [ 17 ] , where nodule-only masks led to implausible nodule placements, our volumetric setting proposes richer conditioning. Spatial and textural cues are encoded by assigning lungs a value of 0.5 and nodules 1–5 (non-solid to solid). Masks are normalized to [0,1], downsampled four times via 3D max pooling, concatenated with the noisy latent, and injected into U-Net cross-attention layers [ 22 , 17 ] . Figure 2: Comparison of unconditional (left) and conditional (right) CT generation using LAND and baseline methods PatchDDM [ 2 ] and WDM [ 6 ] . Mask overlays are shown where applicable. Figure 3: LAND samples conditioned on nodule+lung+texture masks, with increasing texture scores. 3 Experimental Results Datasets and Evaluation Two publicly available datasets were used. LIDC-IDRI [ 1 ] includes 1,010 CT volumes with nodule masks and attribute ratings from four radiologists; for this study, nodule textures scored 1–5 (1: Non-Solid, 2: Non-Solid/Mixed, 3: Part-Solid, 4: Solid/Mixed, 5: Solid) were considered. From NLST [ 20 ] , we selected 881 CT volumes with at least one nodule annotation [ 18 ] and generated the nodule masks using an ad-hoc U-Net. Lung regions in both datasets were segmented with a pre-trained open-source U-Net [ 13 ] . All scans were preprocessed as in [ 6 ] . LIDC-IDRI was used for training, while the NLST subset provided unseen anatomical masks for inference. Evaluation follows the protocol of previous SOTA models [ 6 ] , using Fréchet Inception Distance (FID) [ 10 ] for synthesis quality and MS-SSIM [ 24 ] for sample diversity. FID is computed on 881 real and synthetic CT scans using a ResNet-50 pretrained on 23 medical imaging datasets [ 4 ] ; lower FID indicates closer distributional alignment between real and synthetic samples. MS-SSIM is computed on 10k synthetic pairs, with lower scores indicating higher diversity. Table 1: Comparison of LAND (ours) with SOTA methods. FID values are multiplied by 10^{3} . Unconditional FID↓ FID↓ MS-↓ Mem↓ Conditional FID↓ FID↓ MS-↓ Mem↓ Method (LIDC) (NLST) SSIM (GB) Method (LIDC) (NLST) SSIM (GB) PatchDDM [ 2 ] 317.53 376.4 0.39 19.61 LAND nodule 4.52 5.82 0.3 7.52 WDM [ 6 ] 15.24 32.66 0.27 7.27 LAND nodule+lung 4.48 3.37 0.29 7.52 LAND 5.062 4.76 0.29 7.38 LAND nodule+lung+texture 4.60 3.87 0.29 7.52 Implementation Details Training was performed on a single Nvidia Grid A100-20C (20 GB) GPU. The 3D VAE was trained independently for 100 epochs with AdamW (learning rate 1\times 10^{-4} , batch size 1). The 3D U-Net was trained for 500k steps with AdamW (learning rate 1\times 10^{-5} , batch size 1). The diffusion process used T=1000 timesteps with a linear noise schedule from \beta_{1}=1\times