Reproduce “DiffMix: Diffusion Model-based Data Synthesis for Nuclei Segmentation and Classification in Imbalanced Pathology Image Datasets” on OpenResearch

About this paper

Hyun-Jic Oh, Won-Ki Jeong

DiffMix, developed by researchers at Korea University, introduces a data synthesis method that uses a conditioned diffusion model to create realistic pathology image patches for nuclei segmentation and classification. The approach effectively addresses data imbalance and expands data diversity, significantly improving classification performance, particularly for rare nuclei types, on datasets like GLySAC and CoNSeP.

Key insights

• Denoising Diffusion Probabilistic Models (DDPMs), particularly Semantic Diffusion Models (SDMs), can synthesize highly realistic and semantically consistent pathology images, outperforming prior generative models like GANs and pixel-blending methods.
• The optimal data augmentation strategy for imbalanced datasets depends on the severity of the imbalance; 'enlarging' maps (increasing diversity) are more effective for less imbalanced datasets, while 'balancing' maps (addressing class counts) combined with 'enlarging' maps are crucial for highly imbalanced datasets.
• Controlling the generation process with custom semantic label maps allows for targeted data synthesis to address specific challenges like class imbalance and lack of spatial diversity.

Summary by alphaXiv.