LN-Gen: Rectal Lymph Nodes Generation via Anatomical Features

TL;DR

LN-Gen tackles rectal lymph node segmentation under data scarcity by generating anatomically realistic synthetic LN masks through latent diffusion in Signed Distance Function space, augmented by an anatomical adapter and medical priors. The method employs a three-part pipeline: implicit SDF-based structure synthesis, conditioned diffusion guided by morphology, and priors to ensure plausible background placement and size scaling. In downstream segmentation, LN-Gen synthetic data consistently improves performance across U-Net, nnU-Net, and SwinUNETR, outperforming real-only and DiffTumor-augmented baselines, and ablation confirms the best results arise from combining implicit diffusion with the anatomical adapter and priors. These findings suggest anatomically guided diffusion-based synthesis can alleviate annotation bottlenecks and enhance clinical workflows for rectal cancer management.

Abstract

Accurate segmentation of rectal lymph nodes is crucial for the staging and treatment planning of rectal cancer. However, the complexity of the surrounding anatomical structures and the scarcity of annotated data pose significant challenges. This study introduces a novel lymph node synthesis technique aimed at generating diverse and realistic synthetic rectal lymph node samples to mitigate the reliance on manual annotation. Unlike direct diffusion methods, which often produce masks that are discontinuous and of suboptimal quality, our approach leverages an implicit SDF-based method for mask generation, ensuring the production of continuous, stable, and morphologically diverse masks. Experimental results demonstrate that our synthetic data significantly improves segmentation performance. Our work highlights the potential of diffusion model for accurately synthesizing structurally complex lesions, such as lymph nodes in rectal cancer, alleviating the challenge of limited annotated data in this field and aiding in advancements in rectal cancer diagnosis and treatment.

Figures (4)

• Figure 1: Distribution of the long-axis length of rectal lymph nodes in the training dataset. This figure illustrates the distribution of long-axis lengths for the 903 rectal lymph node samples within the training dataset. The long-axis lengths range from 1.7 mm to 30.0 mm, with the majority of samples measuring between 3 mm and 10 mm. Notably, samples with long-axis lengths exceeding 10 mm are scarce. This distribution reflects the considerable morphological and dimensional variability of rectal lymph nodes. Moreover, it underscores the pronounced imbalance within the dataset, which could potentially pose challenges for the effective training of segmentation models.
• Figure 2: Overview of the proposed LN-Gen. This figure illustrates a three-step methodology for the synthesis and segmentation of rectal lymph node structures. In Step 1, the Anatomic Structure Generation Network ($\phi$) is trained on real anatomical data of rectal lymph nodes to generate synthetic structures. In Step 2, the Conditioned Sample Synthesis Network ($\psi$) is trained on real samples to synthesize rectal lymph node structures with realistic morphology and texture, guided by both morphological and positional information. In Step 3, the anatomical structures generated by $\phi$ are utilized to guide the generation of samples with diverse morphological structures, varying sizes, and realistic textures. These synthetic samples are then incorporated into the original training set to boost the segmentation model’s training.
• Figure 3: The synthetic anatomical structures in SDF presentation. We present both real and synthetic rectal lymph nodes in SDF representation. LN-Gen effectively generates anatomical structures with extensive diversity, remarkable stability, and high quality, ensuring the authenticity of the structures while accurately capturing the detailed surface information of the lymph nodes. Incorrectly generated structures are marked with a red X, and structures with minor flaws are indicated by red arrows.
• Figure 4: Visualization of real and synthetic rectal lymph nodes. We present real and synthetic rectal metastatic lymph nodes, demonstrating that our approach can synthesize lymph nodes of varying sizes and morphologies with high quality and authenticity. The red lines indicate the long-axis length of the rectal lymph nodes.