High-precision label-free virtual H&E staining of 3D holotomography using DAPI-guided conditional diffusion learning

Area of Science:

• Digital pathology
• Computational imaging
• Histopathology

Background:

• Conventional hematoxylin and eosin (H&E) staining is destructive and limited to 2D sections.
• Label-free 3D holotomography (HT) offers a nondestructive imaging alternative.
• Existing virtual staining methods often struggle with preserving nuclear morphology and require complex registration.

Purpose of the Study:

• To develop a practical virtual staining method for generating H&E-like images from label-free 3D HT.
• To preserve nuclear morphology in virtual H&E images.
• To require only HT data at inference, simplifying the workflow.

Main Methods:

• A DAPI-guided conditional diffusion model with shared encoder and two decoder heads (H&E and DAPI) was designed.
• The model was trained using HT as conditional input, predicting diffusion noise for H&E and DAPI targets.
• Weakly paired training was employed, using DAPI only during training for nucleus-centric guidance, leveraging stronger HT-DAPI correspondence over weaker HT-H&E pairs.

Main Results:

• The DAPI-guided diffusion model produced more realistic nuclear morphology and better structural consistency compared to CycleGAN baselines.
• The model achieved lower Frechet Inception Distance (FID) and Kernel Inception Distance (KID) scores on held-out test tiles.
• Specifically, FID was 9.2158 and KID was 0.0091 ± 0.0035, outperforming CycleGAN without DAPI (FID: 14.7447; KID: 0.0434 ± 0.0067).

Conclusions:

• Training-time DAPI guidance effectively enhances virtual H&E generation from label-free HT.
• The method does not require DAPI during inference, simplifying practical application.
• The weakly paired training approach reduces reliance on costly pixel-level registration, supporting scalable, nondestructive digital histopathology.