Reagent-free multiple histological staining via generative AI for rapid pathology diagnosis.

BACKGROUND

Histological special stains (e.g., Masson's trichrome, Elastin van Gieson, periodic acid-Schiff) are typically utilized as a supplementary method to highlight specific tissue structures that may not be visible in the hematoxylin and eosin (H&E) stain. However, conventional pathology work flow suffers from long turn-around waiting times and the use of a large amount of laboratory materials.

METHOD: In this paper, we employ a parameter-efficient fine-tuning method to adapt a general-purpose vision model into a task-specific model, which enables the automatic and simultaneous synthesis of multiple histological stains. By adapting the InstructPix2Pix using Low-Rank Adaptation (LoRA) with a high rank (r = 512), we enable the automatic and simultaneous synthesis of multiple histological stains from label-free autofluorescence (AF) images. Our approach leverages AF images as structural anchors to ensure topological consistency, while text prompts serve as precise semantic controllers for stain selection. Crucially, we introduce the Fréchet Inception Distance using Pathology Language-Image Pre-training embeddings (FID-PLIP) to evaluate generation quality, overcoming the limitations of traditional pixel-wise metrics in capturing diagnostically relevant features.

RESULTS

Experiments demonstrate that our method achieves high structural fidelity and diagnostic consistency, with Pearson correlation coefficients of 0.83 for Masson's Trichrome and 0.92 for Elastin van Gieson against ground truth stains.

CONCLUSION

This reagent-free, time-efficient framework offers a robust alternative for clinical pathological examination, significantly accelerating diagnosis workflows.