ASP-DGRank: Attention-Guided Supervised Patch Ranking with Diffusion-Regularized GAN for Virtual Staining

SRIJITA KHATUA, Manisha Verma, Sudhakar Kumawat

Keywords: Adaptive Attention, Histopathological image translation, Diffusion-GAN, Biomarker, Virtual staining

Abstract: Translating histopathological images across different staining modalities is a challenging task, as maintaining both structural consistency and biomarker integrity is critical for diagnostic reliability. To address this, we present ASP-DGRank, an attention-guided supervised patch-ranking model that integrates adversarial learning, Attention-guided Adaptive Supervised Patch $(A^2SP)$ loss, and diffusion-based regularization. The proposed approach selects patches and adaptively weights them to mitigate the effect of noisy or misaligned regions, while emphasizing diagnostically relevant areas. We evaluate ASP-DGRank using the BCI dataset for the HER2 biomarker and the MIST dataset for four biomarkers (HER2, ER, PR, and Ki67). The method produces virtual stains that are structurally consistent and clinically meaningful. Quantitative analysis with structural and pathology-aware metrics indicates improvements compared to existing approaches. Overall, ASP-DGRank contributes to improving the reliability and interpretability of histopathological image translation.

Primary Subject Area: Image Synthesis

Secondary Subject Area: Generative Models

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 190