Virtual 3D H&E Staining from Phase-contrast Back-illumination Interference Tomography

Alex Baras; Anthony Song; Boyan Zhou; Marisa Morakis; Mayank Golhar; Nicholas Durr

arxiv: 2605.22000 · v1 · pith:ZWZX5UXWnew · submitted 2026-05-21 · 💻 cs.CV · cs.AI

Virtual 3D H&E Staining from Phase-contrast Back-illumination Interference Tomography

Anthony Song , Boyan Zhou , Mayank Golhar , Marisa Morakis , Alex Baras , Nicholas Durr This is my paper

Pith reviewed 2026-05-22 07:37 UTC · model grok-4.3

classification 💻 cs.CV cs.AI

keywords virtual staining3D histopathologyphase-contrast imagingback-illumination interference tomographyH&E stainingnuclei segmentationimage-to-image translationdeep learning

0 comments

The pith

A bidirectional consistency framework turns shift-variant BIT phase volumes into realistic 3D H&E stains while preserving nuclear boundaries.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes a pipeline for converting raw three-dimensional phase-contrast images of unprocessed tissue into H&E-stained volumes that retain accurate nuclear positions and shapes. It does so by releasing the first voxel-paired BIT and fluorescence-nuclei dataset and training a translation network that enforces content consistency at multiple scales in both directions while borrowing style statistics from real H&E images. A sympathetic reader would care because the approach removes the need for physical sectioning and staining, opening the door to rapid volumetric pathology on intact samples. The reported gains in perceptual realism scores and in zero-shot Cellpose segmentation accuracy indicate that the translated volumes are closer to real H&E than earlier methods.

Core claim

The central claim is that bidirectional multiscale content consistency combined with cross-domain style reuse translates BIT volumes that exhibit shift-variant contrast into H&E volumes whose realism metrics exceed prior art and whose downstream 3D nuclei segmentation accuracy and boundary fidelity improve under zero-shot Cellpose evaluation, as shown on the newly introduced HistoBIT3D paired dataset.

What carries the argument

Bidirectional multiscale content consistency with cross-domain style reuse, which simultaneously preserves structural details across scales and transfers realistic H&E appearance from a separate image distribution.

If this is right

Virtual H&E volumes reach state-of-the-art scores on standard realism metrics.
3D nuclei segmentation accuracy rises when the virtual stains are used instead of raw BIT data.
Nuclear boundary preservation improves under the same zero-shot evaluation protocol.
The overall pipeline supplies a scalable route to slide-free volumetric computational histopathology.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same consistency mechanism could be tested on other label-free 3D modalities such as optical coherence tomography or light-sheet microscopy.
If integrated with endoscopic BIT hardware the method might support real-time volumetric assessment inside the body.
The released paired dataset provides a public benchmark that later unsupervised 3D translation algorithms can be measured against.
Consistent 3D virtual stains may reduce inter-observer variability in pathology by supplying uniform volumetric context rather than selected 2D planes.

Load-bearing premise

Paired fluorescence nuclear labels in the HistoBIT3D dataset serve as a reliable quantitative stand-in for true tissue microarchitecture when judging how faithfully virtual H&E outputs preserve nuclear locations and edges.

What would settle it

Zero-shot Cellpose run on the generated virtual H&E volumes yields Dice or boundary-error scores no better than those obtained by running the same segmenter directly on the original BIT volumes.

Figures

Figures reproduced from arXiv: 2605.22000 by Alex Baras, Anthony Song, Boyan Zhou, Marisa Morakis, Mayank Golhar, Nicholas Durr.

**Figure 1.** Figure 1: Overview of the 3D virtual staining pipeline for Back-illumination Interference Tomography (BIT). From left to right: optical setup for volumetric acquisition of voxel-wise paired BIT and fluorescence nuclei data from bulk human tissue; virtual H&E generation using our GAN-based framework (top row); 3D fluorescence nuclei segmentation results (bottom row); and quantitative evaluation via Hausdorff distanc… view at source ↗

**Figure 2.** Figure 2: Architecture showing bidirectional multiscale content consistency for structurally faithful BIT-to-H&E translation and AdaIN-based cross-domain style injection for enhanced realism. 2.2 Vision Transformer CycleGAN Prior work has demonstrated that CycleGAN-based architectures [9], including variants with spatial consistency constraints [16, 17], are effective for virtual H&E staining of label-free phase-co… view at source ↗

**Figure 3.** Figure 3: Row 1: Comparison of virtual H&E results from our method and baseline models against unpaired conventional H&E. Row 2: Zero-shot Cellpose [25] segmentation of virtual H&E images compared with ground-truth nuclei from fluorescence imaging. Green arrows highlight nuclei whose structural content is preserved. Rows 3–4: Diverse tissue samples from the HistoBIT3D dataset, virtually stained from BIT into H&E. T… view at source ↗

read the original abstract

Three-dimensional (3D) histopathology of unprocessed tissues has the potential to transform disease management by enabling volumetric characterization of tissue microarchitecture and in-vivo assessment. Back-illumination Interference Tomography (BIT) is a new phase microscopy technology that provides rapid, non-destructive volumetric imaging of unprocessed tissues. However, translating BIT volumes into clinically interpretable H&E images remains challenging, particularly due to shift-variant contrast and the absence of quantitative validation benchmarks. We introduce HistoBIT3D, the first voxel-wise paired BIT and fluorescence-labeled nuclei dataset, enabling quantitative evaluation of structural preservation in unsupervised virtual staining against ground-truth nuclear distributions. Using this dataset, we present a novel virtual staining framework that translates BIT volumes with shift-variant contrast into realistic H&E volumes by leveraging bidirectional multiscale content consistency and cross-domain style reuse to enhance structural fidelity and perceptual realism. Our method achieves state-of-the-art realism metrics while significantly improving 3D nuclei segmentation accuracy and boundary preservation under zero-shot Cellpose evaluation. Together, these contributions establish a quantitatively validated, structurally faithful, and scalable pipeline for 3D virtual H&E staining, advancing the paradigm of slide-free, volumetric computational histopathology. Our data and code are available at: https://github.com/aasong113/HistoBIT3D_VirtualStaining.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper releases the first paired BIT-fluorescence dataset and adapts virtual staining with bidirectional multiscale consistency for shift-variant 3D contrast, but its quantitative claims lean on nuclear labels as a proxy for full H&E fidelity.

read the letter

The main thing to know is that this work releases HistoBIT3D, the first voxel-wise paired dataset of back-illumination interference tomography volumes with fluorescence nuclear labels, and pairs it with a translation framework that adds bidirectional multiscale content consistency and cross-domain style reuse to handle the shift-variant contrast typical of BIT. They report state-of-the-art realism metrics and clear gains when the outputs are fed into zero-shot Cellpose for 3D nuclei segmentation and boundary preservation. The data and code are public on GitHub, which is a practical plus for anyone who wants to test or extend the approach. The method takes standard image-translation losses and layers on the multiscale bidirectional terms specifically for this modality, which looks like a reasonable engineering step beyond plain 2D virtual staining pipelines. That part is new and directly tied to the technical properties of BIT. The evaluation uses the fluorescence labels as ground truth for structural preservation, and the zero-shot Cellpose check is a useful external probe that avoids training a new segmenter on the virtual outputs. On the softer side, the validation stays centered on nuclear distributions. H&E diagnostic value also comes from eosin contrast in cytoplasm, stroma, and extracellular matrix, so better nuclei segmentation does not automatically confirm that the full microarchitecture is rendered faithfully or that the images would read the same way to a pathologist. The abstract does not give dataset size, exclusion rules, or details on how the realism metrics were computed across the volume, which leaves some uncertainty about how general the gains are. This paper is aimed at groups working on slide-free 3D histopathology, phase microscopy, and computational staining. Readers who need a new public dataset for BIT or who are adapting translation networks to shift-variant 3D data will get the most out of it. The combination of a fresh paired resource and a modality-specific loss design is enough to justify a serious referee, even if the evaluation would benefit from broader checks beyond nuclei. I recommend sending it for peer review with requests to expand the validation metrics and clarify the dataset statistics.

Referee Report

1 major / 0 minor

Summary. The manuscript introduces HistoBIT3D, the first voxel-wise paired dataset of Back-illumination Interference Tomography (BIT) volumes and fluorescence-labeled nuclei from unprocessed tissues. It proposes a virtual staining framework that translates shift-variant BIT volumes into realistic 3D H&E images via bidirectional multiscale content consistency and cross-domain style reuse. The work reports state-of-the-art realism metrics and significantly improved 3D nuclei segmentation accuracy plus boundary preservation under zero-shot Cellpose evaluation against the fluorescence ground truth.

Significance. If the results hold, this would constitute a meaningful advance toward slide-free volumetric computational histopathology by supplying a scalable, quantitatively validated pipeline for generating clinically interpretable 3D H&E images from non-destructive phase-contrast imaging. The public release of the dataset and code is a clear strength that supports reproducibility and community follow-up.

major comments (1)

[Abstract] Abstract: The central claims of structural fidelity and perceptual realism for complete H&E volumes are supported primarily by improved zero-shot Cellpose nuclei segmentation accuracy and boundary preservation evaluated against fluorescence nuclear labels. Because standard H&E renders both nuclear detail (hematoxylin) and cytoplasmic/stromal detail (eosin), nuclear fluorescence alone is an incomplete proxy for the full tissue microarchitecture; additional quantitative checks on non-nuclear features would be required to substantiate the realism claims.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and for highlighting the potential significance of HistoBIT3D toward slide-free volumetric computational histopathology. We address the concern regarding validation of full H&E realism below.

read point-by-point responses

Referee: The central claims of structural fidelity and perceptual realism for complete H&E volumes are supported primarily by improved zero-shot Cellpose nuclei segmentation accuracy and boundary preservation evaluated against fluorescence nuclear labels. Because standard H&E renders both nuclear detail (hematoxylin) and cytoplasmic/stromal detail (eosin), nuclear fluorescence alone is an incomplete proxy for the full tissue microarchitecture; additional quantitative checks on non-nuclear features would be required to substantiate the realism claims.

Authors: We appreciate this observation. The manuscript reports state-of-the-art realism metrics (including FID and perceptual similarity computed over complete generated H&E volumes against real H&E image distributions) that capture overall perceptual quality encompassing both nuclear and non-nuclear (cytoplasmic/stromal) features. The zero-shot Cellpose results provide complementary quantitative evidence of nuclear structural fidelity using the paired fluorescence ground truth. The bidirectional multiscale content consistency and cross-domain style reuse are explicitly designed to maintain full tissue microarchitecture. We agree that distinguishing these evaluation aspects more clearly would strengthen the abstract. We will revise the abstract and add a clarifying paragraph in the results to explicitly separate perceptual realism metrics from nuclear-specific structural validation. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper introduces the new HistoBIT3D dataset of paired BIT volumes and fluorescence nuclear labels, then applies a virtual staining framework based on bidirectional multiscale content consistency and cross-domain style reuse. No equations are shown that equate any output quantity to an input by construction, and no parameters are fitted to a subset then renamed as a prediction. The quantitative claims rest on zero-shot Cellpose evaluation against the external fluorescence ground truth and standard realism metrics, which constitute independent benchmarks rather than self-referential reductions. No load-bearing self-citations, uniqueness theorems from prior author work, or ansatz smuggling appear in the provided text. The central pipeline therefore remains independent of its own fitted values or definitional loops.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Central claims rest on the new paired dataset serving as faithful ground truth and on the unproven effectiveness of the consistency losses for structural fidelity across domains.

free parameters (1)

multiscale consistency loss weights
Hyperparameters balancing content consistency at different scales during network training; not specified in abstract.

axioms (1)

domain assumption Fluorescence nuclear labels accurately represent structural preservation needed for H&E fidelity evaluation.
Invoked when using the paired dataset for quantitative validation of virtual staining outputs.

pith-pipeline@v0.9.0 · 5780 in / 1410 out tokens · 61073 ms · 2026-05-22T07:37:37.800622+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

bidirectional multiscale content consistency loss that aligns hierarchical features across translation cycles... AdaIN-based injection of learned H&E style features
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

zero-shot Cellpose evaluation... 3D nuclei segmentation accuracy and boundary preservation

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages

[1]

Kiemen, A.L., et al.: CODA: quantitative 3D reconstruction of large tissue volumes from serial histology sections.Nature Methods(2022)

work page 2022
[2]

Kathiriya, J.J., et al.: Disrupted boundary organization in cardiovascular mi- croanatomy revealed by three-dimensional histology. (2026)

work page 2026
[3]

Serafin, R., et al.: Non-destructive three-dimensional pathology using open-top light-sheet microscopy. (2023)

work page 2023
[4]

Reddi, S., et al.: A three-dimensional Barrett esophagus atlas for clinical risk as- sessment using open-top light-sheet microscopy. (2023)

work page 2023
[5]

Ching-Roa, V., et al.: Real-time three-dimensional two-photon fluorescence mi- croscopy for skin-biopsy analysis. (2022)

work page 2022
[6]

Yin, C., et al.: Fast label-free three-dimensional virtual histology using optical coherence tomography. (2025)

work page 2025
[7]

Ford, T.N., et al.: Phase-gradient microscopy in thick tissue with oblique back- illumination.Nature Methods(2012)

work page 2012
[8]

McKay, G.N., et al.: Back-illumination interference tomography for label-free vol- umetric tissue imaging. (2025)

work page 2025
[9]

Abraham, T., et al.: Label-free three-dimensional histology with virtual staining. (2023)

work page 2023
[10]

Almagro-Perez, C., et al.: Histology-guided micro-CT and virtual staining for vol- umetric tissue analysis. (2025)

work page 2025
[11]

Song, A.A., et al.: Slide-free tissue histology using back-illumination interference tomography and virtual H&E staining.Optica Biophotonics Congress(2025)

work page 2025
[12]

Park, S., et al.: Three-dimensional virtual staining from holotomography. (2025)

work page 2025
[13]

Rivenson, Y., et al.: Virtual histological staining of unlabelled tissue- autofluorescence images via deep learning.Nature Biomedical Engineering(2019)

work page 2019
[14]

In:ICCV(2017)

Zhu, J.-Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In:ICCV(2017)

work page 2017
[15]

Zhang, Y., et al.: High-throughput virtual histology of thick unstained tissue. (2022)

work page 2022
[16]

Li, X., et al.: Unsupervised content-preserving transformation for optical mi- croscopy virtual staining. (2021)

work page 2021
[17]

You, S., et al.: STABLE: structure-preserving unsupervised virtual staining with feature consistency. (2025)

work page 2025
[18]

In:ICLR(2021)

Dosovitskiy, A., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In:ICLR(2021)

work page 2021
[19]

Hussain, M., et al.: ViT-Stain: vision-transformer-based virtual staining for histopathology. (2026)

work page 2026
[20]

Torbunov, D., Huang, Y., Yu, H., Huang, J., Yoo, S., Lin, M., Viren, B., Ren, Y.: UVCGANv2: an improved cycle-consistent GAN for unpaired image-to-image translation. (2023)

work page 2023
[21]

Chen, X., et al.: Scaling vision transformers for robust visual representation learn- ing. (2022)

work page 2022
[22]

In:ICCV(2017) 10 Anthony A

Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In:ICCV(2017) 10 Anthony A. Song et al

work page 2017
[23]

In:CVPR(2019)

Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In:CVPR(2019)

work page 2019
[24]

Fereidouni, F., et al.: Microscopy with ultraviolet surface excitation for rapid slide- free histology.Nature Biomedical Engineering(2017)

work page 2017
[25]

Stringer, C., Wang, T., Michaelos, M., Pachitariu, M.: Cellpose: a generalist algo- rithm for cellular segmentation.Nature Methods(2021)

work page 2021
[26]

Zhou, B., et al.: u-Segment3D: universal consensus segmentation for three- dimensional microscopy. (2025)

work page 2025
[27]

Wu, J., et al.: CycleDiffusion: diffusion-based unpaired image-to-image translation. (2022)

work page 2022

[1] [1]

Kiemen, A.L., et al.: CODA: quantitative 3D reconstruction of large tissue volumes from serial histology sections.Nature Methods(2022)

work page 2022

[2] [2]

Kathiriya, J.J., et al.: Disrupted boundary organization in cardiovascular mi- croanatomy revealed by three-dimensional histology. (2026)

work page 2026

[3] [3]

Serafin, R., et al.: Non-destructive three-dimensional pathology using open-top light-sheet microscopy. (2023)

work page 2023

[4] [4]

Reddi, S., et al.: A three-dimensional Barrett esophagus atlas for clinical risk as- sessment using open-top light-sheet microscopy. (2023)

work page 2023

[5] [5]

Ching-Roa, V., et al.: Real-time three-dimensional two-photon fluorescence mi- croscopy for skin-biopsy analysis. (2022)

work page 2022

[6] [6]

Yin, C., et al.: Fast label-free three-dimensional virtual histology using optical coherence tomography. (2025)

work page 2025

[7] [7]

Ford, T.N., et al.: Phase-gradient microscopy in thick tissue with oblique back- illumination.Nature Methods(2012)

work page 2012

[8] [8]

McKay, G.N., et al.: Back-illumination interference tomography for label-free vol- umetric tissue imaging. (2025)

work page 2025

[9] [9]

Abraham, T., et al.: Label-free three-dimensional histology with virtual staining. (2023)

work page 2023

[10] [10]

Almagro-Perez, C., et al.: Histology-guided micro-CT and virtual staining for vol- umetric tissue analysis. (2025)

work page 2025

[11] [11]

Song, A.A., et al.: Slide-free tissue histology using back-illumination interference tomography and virtual H&E staining.Optica Biophotonics Congress(2025)

work page 2025

[12] [12]

Park, S., et al.: Three-dimensional virtual staining from holotomography. (2025)

work page 2025

[13] [13]

Rivenson, Y., et al.: Virtual histological staining of unlabelled tissue- autofluorescence images via deep learning.Nature Biomedical Engineering(2019)

work page 2019

[14] [14]

In:ICCV(2017)

Zhu, J.-Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In:ICCV(2017)

work page 2017

[15] [15]

Zhang, Y., et al.: High-throughput virtual histology of thick unstained tissue. (2022)

work page 2022

[16] [16]

Li, X., et al.: Unsupervised content-preserving transformation for optical mi- croscopy virtual staining. (2021)

work page 2021

[17] [17]

You, S., et al.: STABLE: structure-preserving unsupervised virtual staining with feature consistency. (2025)

work page 2025

[18] [18]

In:ICLR(2021)

Dosovitskiy, A., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In:ICLR(2021)

work page 2021

[19] [19]

Hussain, M., et al.: ViT-Stain: vision-transformer-based virtual staining for histopathology. (2026)

work page 2026

[20] [20]

Torbunov, D., Huang, Y., Yu, H., Huang, J., Yoo, S., Lin, M., Viren, B., Ren, Y.: UVCGANv2: an improved cycle-consistent GAN for unpaired image-to-image translation. (2023)

work page 2023

[21] [21]

Chen, X., et al.: Scaling vision transformers for robust visual representation learn- ing. (2022)

work page 2022

[22] [22]

In:ICCV(2017) 10 Anthony A

Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In:ICCV(2017) 10 Anthony A. Song et al

work page 2017

[23] [23]

In:CVPR(2019)

Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In:CVPR(2019)

work page 2019

[24] [24]

Fereidouni, F., et al.: Microscopy with ultraviolet surface excitation for rapid slide- free histology.Nature Biomedical Engineering(2017)

work page 2017

[25] [25]

Stringer, C., Wang, T., Michaelos, M., Pachitariu, M.: Cellpose: a generalist algo- rithm for cellular segmentation.Nature Methods(2021)

work page 2021

[26] [26]

Zhou, B., et al.: u-Segment3D: universal consensus segmentation for three- dimensional microscopy. (2025)

work page 2025

[27] [27]

Wu, J., et al.: CycleDiffusion: diffusion-based unpaired image-to-image translation. (2022)

work page 2022