Advances in Deep Learning-Based Virtual Staining of Pathological Tissues

Junhong Huang; Tingdong Kou; Tianyue He; Cui Huang; Chaoqiang Wu; Junfei Shen

doi:10.3788/LOP242293

Laser & Optoelectronics Progress, Volume. 62, Issue 14, 1400001(2025)

Advances in Deep Learning-Based Virtual Staining of Pathological Tissues

Junhong Huang, Tingdong Kou, Tianyue He, Cui Huang, Chaoqiang Wu, and Junfei Shen^*

D Sensing and Machine Vision Laboratory, College of Electronics and Information Engineering, Sichuan University, Chengdu 610065, Sichuan , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(16)

Fig. 1. Traditional histological staining and virtual staining workflow. (a) Traditional histological staining links (involves tissue extraction, fixation, embedding, sectioning, chemical staining, and microscopic imaging); (b) virtual staining links (involves tissue extraction, fixation, embedding, sectioning, imaging of unstained or stained sections, input into the virtual staining model, and generation of the virtually stained image)

Download full size

Fig. 2. Four stages of virtual staining model construction (including data acquisition, data preprocessing, network design and training, and staining quality evaluation)

Download full size

Fig. 3. Research results of stain transformation. (a) Special stains^{[10-11,14,16,18]}; (b) IHC^[11,23-25]; (c) IF^[29-31]

Download full size

Fig. 4. Research results of guided staining. (a) Fluorescent labeling imaging^[37-38,40]; (b) autofluorescence imaging^[7,43]; (c) nonlinear optical imaging^[50]; (d) photoacoustic imaging^[52-56]

Download full size

Fig. 5. Research results of plain to stain. (a) Reflective-transmissive optical imaging^[60,65]; (b) spectral microimaging^[64]; (c) phase sensitive imaging^[6,68-69,72]

Download full size

Fig. 6. Overall framework under supervised training strategy^[7]

Download full size

Fig. 7. Commonly used generator and discriminator structures under supervised training strategy^[7]

Download full size

Fig. 8. Overall framework under unsupervised training strategy^[18]

Download full size

Fig. 9. Commonly used generator and discriminator structures under unsupervised training^[18]

Download full size

Fig. 10. Objective quantitative analysis (including numerical metrics and external software analysis)^[54]

Download full size

Fig. 11. Subjective qualitative judgment (includes diagnostic consistency assessment and subjective blind selection)^[54]

Download full size

Table 1. Summary of research on stain transformation

View table

Table 1. Summary of research on stain transformation

Kind	Author	Input	Target	Organ	Training method	Evaluation method
Special stains	de Haan et al. ^［10］	HE	PAS， JMS， MT	Kidney	S	S
	Naglah et al.^［11］	HE	MT	Liver	S	O
	Yan et al.^［12］	HE	MT	Liver	U	O， S
	Biswas et al.^［13］	HE	EVG	Pancreas	U	O
	Levy et al.^［14］	HE	Trichrome， SOX10 IHC	Liver， skin	U	O
	Teramoto et al.^［15］	Giemsa	Papanicolaou	Lung	U	S
	Guan et al.^［16］	HE	MAS， PAS， PASM	Kidney， lung	U	O
	Yang et al.^［17］	HE	PAS	Kidney	S	O
IHC	Zhang et al.^［18］	HE	CC10/Ki67 IHC， Oil Red O	Lung， breast， atherosclerotic lesion	S	O
	Liu et al.^［19］	HE	HER2 IHC	Breast	S	O， S
	Xie et al.^［20］	HE	CK8 IHC	Prostate	S	O
	Hong et al.^［21］	HE	CK IHC	Stomach	S	S
	Cetin et al.^［22］	HE	CD8 IHC	T-cell	U	O
	Dubey et al.^［23］	HE	CDX2/CK818 IHC	Colon	U	O
	Lin et al.^［24］	HE	PR IHC	Breast	S	O
	Mercan et al.^［25］	HE	PHH3 IHC	Breast	S	O
	Jackson et al.^［26］	HE	SOX10 IHC	Skin， lymph node	S	O， S
	Trullo et al.^［27］	Ki67/CD3/ CD8 IHC	HE	Bladder， prostate， colon， breast	U	O
IF	Ghahremani et al.^［28］	IHC	Mutiplex IF	Lung， bladder， colon， breast， prostate	S	O
	Wölflein et al.^［29］	Hoechst	CD3/CD8 IF	Kidney	S	O
	Burlingame et al.^［30］	HE	IF	Pancreatic cancer	S	O

Table 2. Summary of research on guided staining

View table

Table 2. Summary of research on guided staining

Kind	Author	Input	Target	Organ	Training method	Evaluation method
Fluorescent labelling imaging	Chen et al.^［38］	Hoechst	HE	Mouse brain	S	O
	Jin et al.^［39］	DAPI+ Rhoda mine B	HE	Oral cancer	S	O
	Zhou et al.^［37］	DAPI + autofluorescence	CD3/PD1/ FOXP3/PD1 IF	Lung cancer	U	O
Autofluorescence imaging	Shi et al.^［40］1	266 nm	HE	Lung	U	O， S
	Zhang et al.^［41］	Digital staining matrix+377 nm +562 nm	HE， Jones， MT	Kidney	S	O
	Bai et al.^［42］	377 nm+465 nm+ 562 nm+628 nm	HER2 IHC	Breast	S	O， S
	Li et al.^［43］	377 nm+ 562 nm	HE	Lung	S	O， S
	Rivenson et al.^［7］	377 nm+628 nm	HE， Jones， MT	Salivary， liver， kidney， thyroid， lung	S	S
Nonlinear optical imaging	Shen et al.^［44］	Non-scanning ultrafast imaging microscope （SHG+2PA+3PA）	HE	Brain， ovarian	U	O
Nonlinear optical imaging	Borhani et al.^［45］	TPEF + FLIM	CD3/CD8 IF	Mouse liver	S	O
Photoacoustic imaging	Cao et al.^［46］	UV-PAM	HE	Skeleton	U
	Kang et al.^［47］	UV-PAM	HE	Mouse brain	U	O， S
	Martell et al.^［48］	UV-PARS+UV scattering microscopy	HE	Breast， prostate	U	O
	Restall et al.^［49］	UV-PARS + scattering microscopy	HE	Breast， stomach， mouse lung	S	O， S
	Boktor et al.^［50］	TA-PARS	HE	Skin	S	S

Table 3. Summary of research on plain to stain

View table

Table 3. Summary of research on plain to stain

Kind	Author	Input	Target	Organ	Training method	Evaluation method
Reflective transmissive optical imaging	Li et al.^［59］	Wide-field microscope	HE， PSR， Orcein	Carotid artery	S	O， S
	Asaf et al.^［60］	Wide-field microscope	HE	Skin	U	O， S
	Li et al.^［61］	RCM	HE	Skin	U	O
Spectral microimaging	Mayerich et al.^［62］	Fourier infrared spectral imaging	HE， MT，IHC	Breast	S	‒
Spectral microimaging	Soltani et al.^［63］	Multi-spectral deep ultraviolet microscopy	HE	Prostate	U	S
Phase-sensitive imaging	Abraham et al.^［64］	QOBM	HE	Mouse liver， human gliomas	U	O， S
	Rivenson et al.^［65］	QPI	HE， Jones， MT	Skin， kidney， liver	S	O
	Min et al.^［66］	Wide-filed DPM	HE	Mouse multi-organs	S	O
	Liu et al.^［6］	OCT	HE	Skin	U	O

Table 4. Summary of data acquisition methods and comparison of advantages and disadvantages

View table

Table 4. Summary of data acquisition methods and comparison of advantages and disadvantages

Method	Input imaging method	Advantage	Disadvantage
Stain transformation	Whole slide scanner	Staining allows visualization of tissue and helps models learn feature mapping	Inability to avoid the limitations of traditional technology
Guided staining	Fluorescent labeling imaging	Ability to visualize structures	Tissue damage， quenching， signal overlap
	Autofluorescence imaging	No introduction of fluorescent labeling， multi-channel images provide rich information	Low signal strength， high noise， signal overlap
	Nonlinear optical imaging	Combining multiple mechanisms with adequate information	Low excitation efficiency， high power laser causing tissue damage
	Photoacoustic imaging	Ability to provide high contrast images of the nucleus and cytoplasm	Low equipment integration， low excitation efficiency
Plain to stain	Reflection transmission optical imaging	Universal equipment， lower tissue damage	Low image resolution， high background noise
	Spectral microimaging	Multiple images enrich information	Slower imaging， complex equipment， noise overlay
	Phase-sensitive imaging	Phase contrast provides informative	Higher requirements for coherent or polarized light

Table 5. Summary of network design and training methods and comparison of advantages and disadvantages

View table

Table 5. Summary of network design and training methods and comparison of advantages and disadvantages

Training method	Generator structure	Advantage	Disadvantage
Supervised	U-Net	Strong multi-scale feature capture and detail preservation， effective integration of low-level and high-level features through skip connections	Complex computation， large number of parameters， high memory consumption， especially for high-resolution images
	U-Net + attention gating	Enhanced attention to key regions for improved detail and color quality， dynamic focus on important areas	Increased parameters， higher training complexity， requires careful tuning of attention-related hyperparameters
	Multi-path/parallel structure	Task separation， multi-scale fusion， increased robustness， improved adaptability in multi-task learning	Complex structure， higher time-consuming， requires larger datasets for effective training
Unsupervised	U-Net	Multi-scale reconstruction， better restoration of details， works well in capturing contextual features	Unmatched data can lead to unstable generation， difficulties in maintaining consistency across domains
	ResNet-based	Mitigates the vanishing gradient problem with deep feature learning， facilitates the capture of high-level features	Neglects local details， leads to image blurring/artifacts， loss of high-frequency information， poorer detail
	CycleGAN variants	Improved cycle consistency， adaptable to specific tasks （e.g.， noise reduction， multi-style transfer）	Complex structure， challenging to tune parameters， requires additional loss functions， difficult training

Tools

Get Citation

Copy Citation Text

Junhong Huang, Tingdong Kou, Tianyue He, Cui Huang, Chaoqiang Wu, Junfei Shen. Advances in Deep Learning-Based Virtual Staining of Pathological Tissues[J]. Laser & Optoelectronics Progress, 2025, 62(14): 1400001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Reviews

Received: Nov. 19, 2024

Accepted: Mar. 13, 2025

Published Online: Jul. 16, 2025

The Author Email: Junfei Shen (shenjunfei@scu.edu.cn)

DOI:10.3788/LOP242293

CSTR:32186.14.LOP242293

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology