Reproduce “LN-Gen: Rectal Lymph Nodes Generation via Anatomical Features” on OpenResearch

About this paper

Weidong Guo, Hantao Zhang, Shouhong Wan, Bingbing Zou, Wanqin Wang, Peiquan Jin

The paper introduces "LN-Gen," a technique for generating realistic 3D synthetic rectal lymph nodes by combining implicit diffusion with an anatomical adapter. Integrating LN-Gen's synthetic data into training datasets consistently improved the Sensitivity and Dice scores of various deep learning segmentation models for rectal lymph nodes by up to 8.53%.

Key insights

• Modeling complex anatomical shapes via implicit SDFs combined with an anatomical adapter effectively overcomes issues of structural discontinuity and low quality typically associated with direct diffusion methods, producing stable and detailed 3D representations.
• Incorporating medical priors, such as realistic contextual positioning and anatomical scaling during synthesis, is crucial for generating clinically plausible and diverse synthetic lesions that genuinely reflect real-world variability.
• Augmenting training datasets with anatomically guided synthetic data significantly enhances the performance, Sensitivity, and Dice scores of various segmentation models, addressing data scarcity and improving generalization for challenging medical tasks.

Summary by alphaXiv.