SpaDiffHis: Sparse-point Guided Diffusion for Histopathology Image Synthesis with Contrastive Learning

Abstract:

We propose a novel framework for generating realistic synthetic histopathology images and their corresponding segmentation masks by fine-tuning a Stable Diffusion model conditioned on labeled point maps. Unlike conventional methods that heavily rely on textual prompts, our approach leverages spatial information from unsupervised nuclei detection and incorporates contrastive learning to enforce the preservation of nucleus structure and class balance. The proposed SpaDiffHis pipeline integrates a contrastive head that aligns the model's latent representations with pre-extracted domain-specific characteristics, ensuring improved fidelity and stain consistency in the synthetic hematoxylin and eosin (H\&E) patches. Extensive experiments demonstrate that our method produces high‐quality images that capture fine histopathological details, while the co-synthesized masks facilitate downstream tasks such as segmentation and classification. We evaluated the framework on the Lizard, PanNuke, and CoNSeP datasets. Experimental results show that our method produces high-quality images with fidelity and diversity close to real slides with Fréchet inception distance (FID) 36.39 and inception score (IS) 2.38 on Lizard, FID 33.53 and IS 3.81 on PanNuke, FID 34.18 and IS 2.18 on CoNSeP respectively, as well as accurate co-synthesized segmentation masks with Dice metrics up to 0.747 for Lizard, 0.816 for PanNuke and 0.857 for CoNSeP. Overall, our proposed diffusion-based synthesis approach generates realistic nuclei morphology and structurally sound tissue characteristics.