Deep Learning for Generating Computational PIN-4 Immunohistochemistry Staining from Prostate Biopsy H&E Images
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Deep Learning for Generating Computational PIN-4 Immunohistochemistry Staining from Prostate Biopsy H&E Images".
Jane: Here's the gist of the paper. Immunohistochemistry (IHC) is frequently used to resolve diagnostically ambiguous prostate cancer biopsy findings on hematoxylin and eosin (H&E)-stained tissue However, a conditional generative adversarial network (cGAN) was trained to synthesize PIN-4 staining patterns directly from native H&E image patches This approach enables direct interpretation of predicted PIN-4 marker patterns in the context of the source prostate H&E architecture, addressing a current spatial limitation of conventional adjacent-section IHC Dataset Construction and Registration The final dataset comprised 172 paired WSIs from 93 patients and 27,298 registered 1024×1024 patch pairs This involved a core-level registration and quality control pipeline for aligning prostate biopsy WSIs before patch extraction, including review and exclusion of misregistered or artifact-containing regions The final cohort included tissue blocks from 93 patients, comprising 172 H&E/PIN-4 WSIs pairs and 27,298 registered patch pairs Model Training Methodology The Pix2Pix model was adapted from a published and clinically benchmarked computational H&E staining model
6: and trained to learn the H&E-to-PIN-4 image patch mapping The generator loss was defined as LcGAN(G, D) = Ex,y
log D(x, y): + Ex,y,z
log(1 − D(x, G(x, z))): , L1(G) = Ex,y,z
∥y − G(x, z)∥1: , LPCC(G) = Ex,y,z
1 − PCC(y, G(x, z)): , and LSSIM(G) = Ex,y,
1 − SSIM(y, G(x, z)): Hyperparameter tuning minimized validation LPIPS across 76 independent evaluations using the Tree-structured Parzen Estimator algorithm The selected model was trained for 60 epochs using λ = 61, γ = 20, and ζ = 4.2 Performance Metrics and Qualitative Assessment Model performance on the held-out test set achieved a mean peak signal-to-noise ratio (PSNR) of 21.88 dB, structural similarity index measure (SSIM) of 0.667, Pearson correlation coefficient (PCC) of 0.684, and learned perceptual image patch similarity (LPIPS) of 0.417 Qualitative review by a board-certified pathologist demonstrated that generated images captured diagnostically relevant PIN-4 staining patterns, including AMACR/racemase expression and basal-cell-associated staining, while preserving spatial correspondence with the source H&E morphology Clinical Interpretation and Future Directions The qualitative findings demonstrated that the model outputs captured several of these relationships, indicating that the model learned morphologically grounded associations rather than color statistics Model errors were observed predominantly in morphologically complex regions such as intraductal carcinoma and high-grade carcinoma, where training examples were limited Future work should expand the dataset with additional cases from multiple institutions to assess generalizability across staining protocols, scanners, tissue processing workflows, and patient populations The study contributes a reproducible workflow for constructing paired, registered H&E/IHC datasets from routine clinical WSIs TABLE I DISTRIBUTION OF THE PAIRED DATASET.
Tom: Patients WSI Patches Count 93 172 27298 TABLE II IMAGE GENERATION METRICS ACROSS 34 VALIDATION AND TEST WSIS WSI Patches LPIPS ↓ SSIM ↑ PCC ↑ PSNR ↑ TABLE III MEAN PER-PATCH RGB CHANNEL VALUES FOR GENERATED AND GROUND TRUTH IMAGES BY ORIGIN WSI Red Channel WSI Gen.
Jane: GT Diff.
Tom: Diff.
Title and authors: Tom: So we're looking at this paper today, titled "Deep Learning for Generating Computational PIN-four Immunohistochemistry Staining from Prostate Biopsy H andE Images" <ref:2606.01871#pg1,Deep Learning for Generating Computational PIN-4 Immunohistochemistry Staining from Prostate Biopsy>. It sounds complicated because it’s trying to solve a problem where pathologists have to look at two separate slides to figure out what cancer is happening.
Jane: That's right, Tom. The paper tackles how we can take an H andE stained image and generate the PIN-four staining pattern directly from it using this conditional generative adversarial network, or cGAN <ref:2606.01871#pg1>. It’s addressing a real issue where you need to see the morphology and the IHC marker side at the same time but they are usually on different sections of tissue.
Lu: The core idea is using the H andE image as a starting point to predict what the PIN-four staining should look like, which lets us interpret those predicted patterns right in context with how that tissue looks under H andE staining <ref:2606.01871#pg1>.
Meng: So, what’s the actual scope of this work? Is it just about generating pretty pictures, or is it actually useful for diagnosis? I need to know if this is something we can actually deploy in a clinical setting without massive overhead.
Tom: Well, they built a final dataset of one hundred seventy-two paired whole slide images from ninety-three patients, which included twenty-seven thousand registered patch pairs, and they used this cGAN to synthesize the PIN-four staining patterns directly from native H andE image patches <ref:2606.01871#pg1>.
Jane: And what they did with that data is important—they created a workflow for constructing these paired, registered H andE and PIN-four datasets from routine clinical whole slide images, which is a big step forward for reproducibility <ref:2606.01871#pg1>.
Lu: They used the Pix2Pix model, adapted from an existing computational H andE staining model to learn that mapping between the two image patches using several loss functions like LcGAN loss and LSSIM loss five.
Meng: That sounds computationally intensive. How much training time did they need for this entire process, and what about the complexity of those generated images? Can we actually trust them to handle difficult cases?
Tom: They trained the model for sixty epochs using specific settings—lambda equals sixty-one gamma equals twenty and zeta equals four point two five. The performance metrics on their held-out test set showed a mean peak signal-to-noise ratio of twenty-one point eight eight decibels and a structural similarity index measure of zero point six six seven Table II <ref:2606.01871#pg3>.
Jane: Those numbers are pretty decent for an image generation task, but the real win here, as the researchers pointed out, is that the generated images captured diagnostically relevant PIN-four staining patterns <ref:2606.01871#pg2>.
Lu: They specifically mentioned capturing things like AMACR and racemase expression and basal-cell associated staining while keeping it spatially matched to the source H andE morphology <ref:2606.01871#pg3>. This suggests the model learned morphologically grounded associations rather than just relying on color statistics <ref:2606.01871#pg3>.
Title and authors: Meng: I mean, capturing those specific marker expressions is good, but what about areas where the training data was scarce? Like high-grade carcinoma or intraductal carcinoma? That's where models often fail when they haven't seen enough examples <ref:2606.01871#pg3>.
Tom: They did note that model errors were predominantly in those morphologically complex regions, like intraductal carcinoma and high-grade carcinoma, because the training examples there were limited <ref:2606.01871#pg3>.
Jane: It sounds like the system isn't perfect everywhere yet, but it’s still showing promise for capturing these relationships when it does perform well <ref:2606.01871#pg2>. What do you think this means for a pathologist who is dealing with ambiguous biopsy findings?
Lu: It suggests that this tool could serve as an adjunctive colocalization tool to review predicted PIN-four marker patterns spatially against the source H andE architecture three. It aims to fix the limitation where conventional adjacent-section IHC can't guarantee that the glands assessed on one section are actually the same ones seen on another <ref:2606.01871#pg1>.
Meng: If it acts as a colocalization tool, how much time does that save? Is this something you could realistically integrate into a standard pathology workflow, or is it just interesting research for now?
Tom: The goal is to enable direct interpretation of predicted PIN-four marker patterns within the context of the source H andE architecture <ref:2606.01871#pg1>. It moves us away from having to jump between slides and trying to manually match features <ref:2606.01871#pg2>.
Jane: So, for someone listening who isn't a pathologist, the big picture is that this AI helps bridge the gap between what you see under one stain and what you expect to see under another marker, by synthesizing that expected pattern right there on the original image <ref:2606.01871#pg2>.
Lu: The training methodology itself is also an interesting piece of work; they used hyperparameter tuning via the Tree-structured Parzen Estimator algorithm to minimize validation LPIPS across seventy-six independent evaluations five.
Meng: That sounds like a lot of tuning. When you move from a research setting to production, how do you handle that level of hyperparameter sensitivity? What breaks at scale when things get messy?
Tom: They settled on training for sixty epochs using lambda equals six one and gamma equals twenty, and zeta equals four point two five. The key here is that they were able to tune it down to a specific set of values that minimized the LPIPS loss five.
Jane: So, for us, what this means is we have a reproducible workflow for building these paired datasets from routine clinical whole slide images, which helps everyone in the field build on this work <ref:2606.01871#pg1>. It gives us a better foundation than just looking at isolated patches.
Lu: And the data construction itself is quite rigorous; they had to do core-level registration and quality control before patch extraction, including reviewing and excluding misregistered or artifact-containing regions <ref:2606.01871#pg1>.
Meng: That pre-processing step sounds crucial for reliability. If you mess up the alignment at that stage, the whole model will be learning bad associations, right?
Title and authors: Tom: Exactly. The final cohort included tissue blocks from ninety-three patients, with one hundred seventy-two H andE and PIN-four WSIs pairs and twenty seven thousand two hundred ninety eight registered patch pairs Table I <ref:2606.01871#pg1>. That large number of paired data is what makes the training possible.
Jane: It's a huge undertaking to get that level of paired registration done for so many samples, but it sets a new standard for what's required when you want to compare two different staining methods spatially <ref:2606.01871#pg2>.
Lu: Looking ahead, the paper points toward expanding this dataset significantly by including additional cases from multiple institutions to test how well the model generalizes across different staining protocols and scanners Future Directions.
Meng: Generalization across different scanners is a major concern for any deployment. If you train it on one scanner's data, it might not work on another, and that’s a huge operational headache Future Directions.
Tom: And the study also contributed a reproducible workflow for constructing these paired registered H andE/IHC datasets from routine clinical whole slide images Future Directions. That part is really valuable for anyone trying to standardize this type of computational staining downstream.
Jane: So, to wrap up this paper, the main thing we learned is that this system captures diagnostically relevant PIN-four staining patterns while preserving spatial correspondence with the source H andE morphology <ref:2606.01871#pg3>. It’s an improvement because it shows it learns associations grounded in morphology, not just random colors <ref:2606.01871#pg3>.
Lu: But we have to keep the caveats in mind; the paper admits that model errors are still higher in morphologically complex regions like high-grade carcinoma <ref:2606.01871#pg3>. That's a clear limitation we need to address with future work.
Meng: So, even with those limitations, it's a tool that can help pathologists review predicted patterns in spatial context with the source H andE architecture, which is better than just looking at adjacent sections three <ref:2606.01871#pg2>. But we still have those tricky areas where it needs more training examples.
Tom: It’s an interesting approach because it doesn't just spit out a stain; it synthesizes the expected staining pattern based on the underlying tissue structure <ref:2606.01871#pg1>.
Jane: So, to summarize this paper, we have a conditional generative adversarial network trained on paired H andE and PIN-four images that generates synthetic PIN-four patterns directly from H andE patches <ref:2606.01871#pg2>. It's a tool for direct morphologic comparison and interpretation <ref:2606.01871#pg3>.
Lu: And the method itself involves adapting the Pix2Pix model using specific loss functions to learn this image-to-image mapping five. The reliance on LcGAN loss, L1 loss, LPCC loss, and SSIM loss shows a comprehensive effort to capture both perceptual similarity and structural fidelity five.
Meng: From a deployment standpoint, we'd be watching how robust those performance metrics hold up when the input data shifts—like if we move from one scanner to another—because that’s where the real friction is for us Future Directions.
Tom: Right. The paper provides a solid foundation by showing that this approach can capture features like AMACR and racemase expression, which are key markers in these cases <ref:2606.01871#pg3>.
Title and authors: Jane: It’s about building a reproducible workflow for creating these paired datasets from routine clinical whole slide images, which is a big step forward for the entire field <ref:2606.01871#pg1>.
Lu: And the potential to use this as an adjunctive tool to review predicted PIN-four marker patterns in spatial context with the source H andE architecture is what makes this really interesting from a representation learning perspective three.
Meng: It's definitely not a replacement for clinical decision making yet; it’s an aid, and we need to see those results hold up when we put it through real traffic at scale Meng.
Tom: So, the takeaway here is that by learning from native H andE patches, this AI can synthesize PIN-four staining patterns directly, which means we get a more direct way to interpret the predicted marker patterns in the context of the original tissue structure <ref:2606.01871#pg1>.
Jane: We're leaving this paper with an understanding that while it captures many relevant features, especially in areas where training is abundant, we still need to focus on expanding the dataset and improving robustness in those more complex regions <ref:2606.01871#pg3>.
Lu: And the next step clearly involves expanding the dataset with additional cases from multiple institutions to assess generalizability across staining protocols and patient populations Future Directions.
Meng: I'm just hoping that when they do that, the performance metrics stay within a reasonable margin of error for deployment, because right now, we see errors in complex areas <ref:2606.01871#pg3>.
Tom: So, to wrap up on "Deep Learning for Generating Computational PIN-four Immunohistochemistry Staining from Prostate Biopsy H andE Images," the paper provides a method using a cGAN trained on paired data to synthesize PIN-four staining patterns directly from native H andE patches <ref:2606.01871#pg1>. It’s an improvement because it shows it learns morphologically grounded associations, which is valuable for interpretation <ref:2606.01871#pg3>.
Jane: This work contributes a reproducible workflow for constructing these paired registered H andE and PIN-four datasets from routine clinical whole slide images <ref:2606.01871#pg1>. It sets the stage for more direct spatial comparison between morphology and IHC signals <ref:2606.01871#pg2>.
Lu: The study also shows that this system can serve as a potential adjunctive colocalization tool to review predicted PIN-four marker patterns in spatial context with the source H andE architecture three. It’s a way to address the limitation where conventional adjacent-section IHC cannot guarantee that the glands assessed on one section are actually the same glands seen on another <ref:2606.01871#pg1>.
Meng: So, it’s an interesting advancement in how we can use generative models to create synthetic diagnostic information from existing clinical data, but we still have that caveat about morphologically complex regions causing errors <ref:2606.01871#pg3>.
Tom: Right. It's a solid piece of work because it establishes the necessary infrastructure for training these models on paired, registered whole slide images <ref:2606.01871#pg1>. We’re looking forward to seeing how they tackle those scaling and generalization issues next Future Directions.
The paper's summary: Tom: So, to wrap up what we just covered, this paper is about using a conditional generative adversarial network to take an H andE stained image patch and generate the expected PIN four staining pattern directly from it.
Jane: Exactly Tom, so it’s not just guessing; the AI is learning the relationship between how tissue looks under one stain and what the marker staining should look like in another place on that same tissue.
Tom: The big deal is that this lets us interpret those predicted patterns right alongside the original H andE structure, which used to be a real problem because you couldn't guarantee they were from the same biopsy <ref:2606.01871#pg1>.
Jane: Right, so it’s building a way for AI to act like a spatial translator between morphology and molecular markers. It learns things like specific expressions, such as AMACR and racemase staining, while keeping that spatial relationship intact with the original H andE layout <ref:2606.01871#pg3>.
Tom: The results on the test set showed a structural similarity index measure of zero point six six seven and a mean peak signal-to-noise ratio of twenty-one point eight eight decibels, which means the generated images look structurally close to the real ones <Table II>.
Jane: But there’s a catch, Tom. The authors admit that their model struggles more in complex areas like high grade or intraductal carcinoma because they simply didn't have enough examples for those specific cases <ref:2606.01871#pg3>.
Tom: That makes sense from an engineering standpoint; you can only train a model as well as the data you feed it, so it’s not perfect in every single scenario yet.
Jane: But what this means for the field is that we finally have a reproducible workflow for creating these paired, registered H andE and PIN four datasets from routine clinical whole slide images <ref:2606.01871#pg1>.
Tom: That infrastructure part is huge; being able to create this kind of paired data reliably sets a new standard for how we build these computational tools <ref:2606.01871#pg1>.
Jane: And because it learns the structure rather than just random colors, this system can serve as a real adjunctive tool for pathologists to review predicted marker patterns in spatial context with the source H andE architecture <three>.
Tom: So we’re moving toward a system where you don't just look at two separate slides; you look at one image and get the predicted staining pattern right there on top of it.
Jane: And that’s what makes this work so interesting—it’s about getting direct, morphologically grounded associations instead of relying on color statistics alone <ref:2606.01871#pg3>.
Tom: It’s a solid step forward because we have the mechanism to synthesize that expected staining pattern based on the underlying tissue structure <ref:2606.01871#pg1>.
Jane: However, we still need to tackle those limitations in complex regions and expand the dataset across different institutions so this tool actually works well everywhere <Future Directions>.
The paper's improvements: Tom: So, to recap, we just talked about how this AI learns to synthesize PIN four staining patterns directly from H andE patches using a Pix two Pix model <ref:2606.01871#pg5>.
Jane: Right, and now we're looking at what the authors suggest they can do next with this kind of work. What are the specific improvements they propose?
Tom: The main thing is expanding the training data significantly, by including additional cases from multiple institutions <Future Directions>. They want to test how well this model works when it sees different staining protocols and different scanners <Future Directions>.
Jane: That makes sense, because right now, if you train a model on one scanner's data, it might totally fail when you try to use it on a different machine. It needs that broad experience.
Tom: And they’re also focusing on building a better way to construct the dataset itself. They want to establish a reproducible workflow for creating these paired registered H andE and PIN four datasets from routine clinical whole slide images <Future Directions>.
Jane: That is really important, Tom, because having a reliable way to build this data means we can actually compare things properly across the whole medical community. It moves this from a single study to something everyone can use <ref:2606.01871#pg1>.
Tom: And they want to make sure the model handles those really tricky areas better, especially high grade and intraductal carcinoma where the training examples were limited <Future Directions>. They’re aiming for better performance there.
Jane: So they are trying to fix the weak spots in their current performance, which is smart because a tool that fails in complex cases isn't very useful in a clinical setting <ref:2606.01871#pg3>.
Tom: Exactly, and they’re also looking into how we can make sure the model learns those deep structural associations rather than just superficial color patterns <ref:2606.01871#pg3>.
Jane: It's about making sure the AI understands *why* a pattern looks that way based on the tissue structure, not just memorizing pixel values. That’s a deeper level of learning and it’s what makes it useful for diagnosis <ref:2606.01871#pg3>.
Tom: So they are pushing for more robust data and better handling of those hard cases while making the whole process repeatable. It sounds like they're moving from a proof of concept to something that’s actually ready for real-world use, if you can get past those complexity hurdles <Future Directions>.
Conclusion: Tom: So, to wrap up on "Deep Learning for Generating Computational PIN four Immunohistochemistry Staining from Prostate Biopsy H andE Images," this paper shows we can use a conditional generative adversarial network to generate synthetic PIN four patterns directly from H andE patches <ref:2606.01871#pg1>.
Jane: It’s about using the tissue's original shape to predict what the marker staining should look like, which helps us interpret those predicted patterns right in context with the source H andE architecture <ref:2606.01871#pg3>.
Tom: The big takeaway is that this approach moves us away from having to jump between slides and try to manually match features for adjacent section IHC, which is a major spatial limitation <ref:2606.01871#pg2>.
Jane: And it provides a reproducible workflow for constructing these paired registered H andE and PIN four datasets from routine clinical whole slide images <ref:2606.01871#pg1>. That infrastructure is key for the whole field.
Tom: I agree, being able to create this kind of paired data reliably sets a new standard for how we build these computational tools <ref:2606.01871#pg5>.
Jane: But we still have those limitations in complex regions like high grade carcinoma where the training examples were limited <ref:2606.01871#pg3>. That means it’s not perfect everywhere yet.
Tom: True, but the way they learn these morphologically grounded associations is what makes it more useful than just a simple color translation <ref:2606.01871#pg3>.
Jane: It’s an aid right now, serving as a potential adjunctive tool for pathologists to review predicted marker patterns in spatial context with the source H andE architecture <ref:2606.01871#pg3>.
Tom: Right. We need to keep pushing that dataset expansion and generalization across different scanners so this thing actually works everywhere, not just on their test set <Future Directions>.
Jane: I think the next big step for us is seeing how robust these results hold up when we put them through real traffic at scale, because those performance metrics are still conditional on what they trained it on <Meng].
Lu: From a representation learning standpoint, this is interesting because it shows how to map functional descriptions, like H andE structure, to another modality like PIN four staining through a unified framework <ref:2606.01871#pg5>.
Meng: I just want to know what the operational cost looks like for deploying something that relies on this level of image registration and training data quality <Future Directions>.
Lalam: I am sure of my answer; the paper is very sure that by learning from native H andE patches, this AI can synthesize PIN four staining patterns directly, which means we get a more direct way to interpret the predicted marker patterns in the context of the original tissue structure <ref:2606.01871#pg1>.
University of California, Irvine
cs.CV
Submitted: 2026-06-01
Updated: 2026-06-01
Journal ref: 2026 25th International Conference on Machine Learning and Applications (ICMLA), IEEE, 2026
DOI: 10.1109/ICMLA69890.2026.00166
Code: https://github.com/mint-lab/roi
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 78/100
The gist: Immunohistochemistry (IHC) is frequently used to resolve diagnostically ambiguous prostate cancer biopsy findings on hematoxylin and eosin (H&E)-stained tissue However…
Key concepts
- Conditional Generative Adversarial Network (cGAN)
- A type of deep learning model used here to create new images. It has two parts: a generator that creates synthetic PIN-4 stains from H&E images, and a discriminator that tries to tell the difference between real and generated stains. This setup learns to map the source H&E image directly to the desired PIN-4 staining pattern.
- Pix2Pix Model
- A specific type of Pix2Pix model adapted for this task. It is trained using several loss functions, including adversarial loss (LcGAN), L1 loss, Pearson correlation coefficient (LPCC), and structural similarity index measure (LSSIM). These losses guide the generator to produce synthetic images that look both realistic and structurally similar to the ground truth PIN-4 stains.
- Spatial Registration
- The process of aligning two images—the source H&E biopsy and the target PIN-4 stain—so they correspond perfectly in space. This is crucial because the model learns to predict staining patterns based on their correct spatial location relative to the underlying tissue architecture, ensuring that predicted markers match where they should be.
- PIN-4 Staining Patterns
- These are specific visual markers used in immunohistochemistry (IHC) to identify certain prostate cancer characteristics. The model learns to generate these patterns, such as AMACR/racemase expression and basal-cell staining, based on the morphological features present in the original H&E tissue patch.
Terminology
Summary
Immunohistochemistry (IHC) is frequently used to resolve diagnostically ambiguous prostate cancer biopsy findings on hematoxylin and eosin (H&E)-stained tissue However, a conditional generative adversarial network (cGAN) was trained to synthesize PIN-4 staining patterns directly from native H&E image patches This approach enables direct interpretation of predicted PIN-4 marker patterns in the context of the source prostate H&E architecture, addressing a current spatial limitation of conventional adjacent-section IHC
Dataset Construction and Registration
The final dataset comprised 172 paired WSIs from 93 patients and 27,298 registered 1024×1024 patch pairs This involved a core-level registration and quality control pipeline for aligning prostate biopsy WSIs before patch extraction, including review and exclusion of misregistered or artifact-containing regions The final cohort included tissue blocks from 93 patients, comprising 172 H&E/PIN-4 WSIs pairs and 27,298 registered patch pairs
Model Training Methodology
The Pix2Pix model was adapted from a published and clinically benchmarked computational H&E staining model [6] and trained to learn the H&E-to-PIN-4 image patch mapping The generator loss was defined as LcGAN(G, D) = Ex,y[log D(x, y)] + Ex,y,z[log(1 − D(x, G(x, z)))], L1(G) = Ex,y,z[∥y − G(x, z)∥1], LPCC(G) = Ex,y,z[1 − PCC(y, G(x, z))], and LSSIM(G) = Ex,y,[1 − SSIM(y, G(x, z))] Hyperparameter tuning minimized validation LPIPS across 76 independent evaluations using the Tree-structured Parzen Estimator algorithm The selected model was trained for 60 epochs using λ = 61, γ = 20, and ζ = 4.2
Performance Metrics and Qualitative Assessment
Model performance on the held-out test set achieved a mean peak signal-to-noise ratio (PSNR) of 21.88 dB, structural similarity index measure (SSIM) of 0. p. 0 The
model was evaluated on an independent held-out test set of 1,814
patch pairs from 17 WSIs, achieving a mean peak signal-to-noise
ratio (PSNR) of 21. 88 dB, structural similarity index measure
(SSIM) of 0. 88 dB, structural similarity index measure
(SSIM) of 0. 667, Pearson correlation coefficient (PCC) of 0. 667, Pearson correlation coefficient (PCC) of 0. 684,
and learned perceptual image patch similarity (LPIPS) of 0.Where the paper says it
Where the paper says it
Where the paper says it
Clinical Interpretation and Future Directions
The qualitative findings demonstrated that the model outputs captured several of these relationships, indicating that the model learned morphologically grounded associations rather than color statistics Model errors were observed predominantly in morphologically complex regions such as intraductal carcinoma and high-grade carcinoma, where training examples were limited Future work should expand the dataset with additional cases from multiple institutions to assess generalizability across staining protocols, scanners, tissue processing workflows, and patient populations The study contributes a reproducible workflow for constructing paired, registered H&E/IHC datasets from routine clinical WSIs
TABLE I
DISTRIBUTION OF THE PAIRED DATASET.
Patients WSI Patches
Count 93 172 27298
TABLE II
IMAGE GENERATION METRICS ACROSS 34 VALIDATION AND TEST WSIS
WSI Patches LPIPS ↓ SSIM ↑ PCC ↑ PSNR ↑
TABLE III
MEAN PER-PATCH RGB CHANNEL VALUES FOR GENERATED AND GROUND TRUTH IMAGES BY ORIGIN WSI
Red Channel
WSI Gen. GT Diff. |Diff.|
1 191.9±16.7 188.4±19.5 3.5±7.1 6.0±5.1
Green Channel
WSI Gen. GT Diff.|Diff.|
2073 ±6, Mean channel values (± SD) for generated (Gen. p. 7 TABLE III
MEAN PER-PATCH RGB CHANNEL VALUES FOR GENERATED AND GROUND TRUTH IMAGES BY ORIGIN WSI
Red Channel
WSI Gen. 1
Mean 207.Where the paper says it
Blue Channel
Mean channel values (± SD) for generated (Gen.) and ground truth (GT) patches by origin WSI
--- Page 1 ---
Deep Learning for Generating Computational PIN-4 Immunohistochemistry Staining from Prostate Biopsy H&E Images
Vietbao Tran
Pratik Shah
--- Page 2 ---
Visual domains. Conditional generative adversarial neural networks (cGANs) have been applied to generate H&E-like images from unstained tissue and to destain chemically stained WSIs Spatial alignment between H&E and IHC WSIs is necessary for paired computational staining, patch-level learning, and direct morphologic comparison
--- Page 3 ---
B. Computational Staining and Image Registration
Global methods estimate a single transformation across the image, such as rigid, similarity, affine, or homography transformations Deformable registration methods address these local differences by estimating a spatially varying displacement field
--- Page 4 ---
D. Registration Quality Metrics
The Dice index between paired tissue masks was used as the primary objective registration quality metric Normalized PCC, PSNR, and SSIM were also calculated on grayscaleconverted core pairs as secondary registration metrics
--- Page 5 ---
G. Pix2Pix Model
The loss functions were defined as: LcGAN(G, D) = Ex,y[log D(x, y)] + Ex,y,z[log(1 − D(x, G(x, z)))], L1(G) = Ex,y,z[∥y − G(x, z)∥1], LPCC(G) = Ex,y,[1 − PCC(y, G(x, z))], and LSSIM(G) = Ex,y,[1 − SSIM(y, G(x, z))]
--- Page 6 ---
Fig. 6. Representative examples of generated PIN-4 patches displayed as 1024 × 1024 pixel patches at 20× magnification. p. 5 Representative examples of generated PIN-4 patches displayed as
1024 × 1024 pixel patches at 20× magnification. Each column shows
one example: (a) source H&E patch, (b) generated PIN-4 patch, and (c)
corresponding ground truth PIN-4 patch. Representative examples of generated PIN-4 patches displayed as
1024 × 1024 pixel patches at 20× magnification. Each column shows
one example: (a) source H&E patch, (b) generated PIN-4 patch, and (c)
corresponding ground truth PIN-4 patch.Where the paper says it
Where the paper says it
--- Page 7 ---
TABLE II
IMAGE GENERATION METRICS ACROSS 34 VALIDATION AND TEST WSIS
WSI Patches LPIPS ↓ SSIM ↑ PCC ↑ PSNR ↑
--- Page 8 ---
TABLE III
MEAN PER-PATCH RGB CHANNEL VALUES FOR GENERATED AND GROUND TRUTH IMAGES BY ORIGIN WSI
Red Channel
Mean 207.8±5.8 209.0±7.9-1.2±3.6 4,0±2,2
Green Channel
Mean 212.8±5.1 213, Mean channel values (± SD) for generated (Gen. p. 7 TABLE III
MEAN PER-PATCH RGB CHANNEL VALUES FOR GENERATED AND GROUND TRUTH IMAGES BY ORIGIN WSI
Red Channel
WSI Gen. 1
Mean 207.Where the paper says it
Blue Channel
Mean 212.4±5, Median 3,4
--- Page 9 ---
ACKNOWLEDGMENT & CODE
Dr. Alarice Lowe performed pathology interpretations of model outputs and ground truth images Code available at: https://github. p. 5 Alarice Lowe performed pathology interpretations of model
outputs and ground truth images. Code available at: https://github. Alarice Lowe performed pathology interpretations of model
outputs and ground truth images. Code available at: https://github.Where the paper says it
Where the paper says it
--- Page 10 ---
REFERENCES
[1] R. L. Siegel, T. B. Kratzer, N. p. 5 com/Dr-Pratik-Shah-UCI/genAI V IHC
REFERENCES
[1] R. Kratzer, N. Sung, and A. Jemal, “Cancer
statistics, 2026,” CA: A Cancer Journal for Clinicians, vol.Where the paper says it
Where the paper says it
[2] P. A. Humphrey, “Histopathology of Prostate Cancer,” Cold Spring Harbor Perspectives in Medicine, vol. p. 5 e70043, 2026. Humphrey, “Histopathology of Prostate Cancer,” Cold Spring
Harbor Perspectives in Medicine, vol.Where the paper says it
[3] B. Delahunt, R J Miller, J R Srigley et al. p. 6 Delahunt, R. Srigley, A.Where the paper says it
[4] B Sabata, B Babenko, R Monroe et al., “Automated Analysis of PIN-4 Stained Prostate Needle Biopsies,” in Prostate Cancer Imaging. p. 6 Monroe, and C. Srinivas, “Automated
Analysis of PIN-4 Stained Prostate Needle Biopsies,” in Prostate Cancer Imaging. Computer-Aided Diagnosis, Prognosis, and Intervention,
A. Madabhushi, J. Madabhushi, J. Dowling, P. Dowling, P. Fenster, P. Fenster, P. Abolmaesumi et al.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[5] V Ravery et al., “The 20-core prostate biopsy protocol–a new gold standard? p. 6 Ravery et al. , “The 20-core prostate biopsy protocol–a new gold
standard?Where the paper says it
[6] A Rana, A Lowe, M Lithgow, K Horback et al., “Use of Deep Learning to Develop and Analyze Computational Hematoxylin and Eosin Staining of Prostate Core Biopsy Images for Tumor Diagnosis,” JAMA network open vol 3 no 5 p e205111
[7] M T Martell, N J M Haven, B D Cikaluk et al., “Deep learning-enabled realistic virtual histology with ultraviolet photoacoustic remote sensing microscopy,” Nature Communications vol 14 no 1 p 5967
[8] J Li, J Garfinkel, X Zhang, D Wu, Y Zhang et al., “Biopsy-free in vivo virtual histology of skin using deep learning,” Light: Science & Applications vol 10 no 1 p 233
[9] Y Li, N Pillar, J Li, T Liu, D Wu et al., “Virtual histological staining of unlabeled autopsy tissue,” Nature Communications vol 15 no 1 p 1684
[10] M Chen, N J Tustison, R Jena, and J C Gee, “Image Registration: Fundamentals and Recent Advances Based on Deep Learning,” in Machine Learning for Brain Disorders, O Colliot, Ed. New York, NY: Springer US 2023 pp 435–458
[11] K Faust, M K Lee, A Dent, C Fiala et al., “Integrating morphologic and molecular histopathological features through whole slide image registration and deep learning,” Neuro-Oncology Advances vol 4 no 1 p vdac001
[12] A Shafique, M Babaie, M Sajadi, A Batten, S Skdar et al., “Automatic Multi-Stain Registration of Whole Slide Images in Histopathology,” in 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), 2021 pp 3622–3625
[13] G Lippolis, A Edsjo, L Helczynski, A Bjartell, and N C Overgaard “Automatic registration of multi-modal microscopy images for integrative analysis of prostate tissue sections,” BMC Cancer vol 13 no 1 p 408
[14] S Klein, M Staring, K Murphy, M A Viergever, and J P W Pluim “A Toolbox for Intensity-Based Medical Image Registration,” IEEE Transactions on Medical Imaging vol 29 no 1 pp 196–205
[15] J Li, J Dong, S Huang, X Li, J Jiang et al., “Virtual Immunohistochemistry Staining for Histological Images Assisted by Weaklysupervised Learning,” in 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024 pp 11 259–11 268
[16] Q Peng, W Lin, Y Hu, A Bao, C Lian et al., “Advancing H&Eto-IHC Virtual Staining with Task-Specific Domain Knowledge for HER2 Scoring,” in Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, M G Linguraru, Q Dou, A Feragen, S Giannarou, B Glocker et al. p. 6 , “Advancing H&Eto-IHC Virtual Staining with Task-Specific Domain Knowledge for
HER2 Scoring,” in Medical Image Computing and Computer Assisted
Intervention – MICCAI 2024, M. Linguraru, Q.Where the paper says it
[17] P F Wong, C McNeil, Y Wang, J Paparian, C Santori et al., “ClinicalGrade Validation of an Autofluorescence Virtual Staining System With Human Experts and a Deep Learning System for Prostate Cancer,” Modern Pathology vol 37 no 11 p 100573
[18] J Pocock, S Graham, Q D Vu, M Jahanifar, S Deshpande et al., “TIAToolbox as an end-to-end library for advanced tissue image analytics,” Communications Medicine vol 2 no 1 p 120
[19] N C Quy and S Choi, “mint-lab/roi picker,” 2025. [Online]. Available: https://github. p. 5 Alarice Lowe performed pathology interpretations of model
outputs and ground truth images. Code available at: https://github.Where the paper says it
[20] C Gatenbee and A R Anderson, “VALIS: Virtual alignment of pathology image series for multi-gigapixel whole slide images,” 2025. [Online]. Available: https://www. p. 6 [21] “Sedeen Viewer | Pathcore,” accessed: 2026-01-03. Available:
https://www.Where the paper says it
[21] “Sedeen Viewer Pathcore,” accessed: 2026-01-03
[24] Z Wang, A Bovik, H Sheikh, and E Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE Transactions on Image Processing vol 13 no 4 pp 600–612
[25] R Liaw, E Liang, R Nishihara, P Moritz, J E Gonzalez et al., “Tune: A Research Platform for Distributed Model Selection and Training,” 2018, arXiv:1807. p. 6 Gonzalez et al. , “Tune:
A Research Platform for Distributed Model Selection and Training,”
2018, arXiv:1807. , “Tune:
A Research Platform for Distributed Model Selection and Training,”
2018, arXiv:1807. 05118 [cs].Where the paper says it
Where the paper says it
[26] J Bergstra, D Yamins, and D Cox, “Making a Science of Model Search: Hyperparameter Optimization in Hundreds of Dimensions for Vision Architectures,” in Proceedings of the 30th International Conference on Machine Learning. p. 6 Yamins, and D. Cox, “Making a Science of Model Search:
Hyperparameter Optimization in Hundreds of Dimensions for Vision
Architectures,” in Proceedings of the 30th International Conference on
Machine Learning.Where the paper says it
[27] R Zhang, P Isola, A A Efros, E Shechtman, and O Wang “The Unreasonable Effectiveness of Deep Features as a Perceptual Metric,” 2018, arXiv:1801.03924 [cs]. p. 6 Wang, “The
Unreasonable Effectiveness of Deep Features as a Perceptual Metric,”
2018, arXiv:1801. 03924 [cs].Where the paper says it
[28] D P Kingma and J Ba, “Adam: A Method for Stochastic Optimization,” 2017, arXiv:1412. p. 6 Kingma and J. Ba, “Adam: A Method for Stochastic Optimization,”
2017, arXiv:1412.Where the paper says it
[29] S J Reddi, S Kale, and S Kumar, “On the Convergence of Adam and Beyond,” 2019, arXiv:1904. p. 6 Kale, and S. Kumar, “On the Convergence of Adam and
Beyond,” 2019, arXiv:1904. Kumar, “On the Convergence of Adam and
Beyond,” 2019, arXiv:1904. 09237 [cs].Where the paper says it
Where the paper says it
[30] R Zhang, P Isola, A A Efros, E Shechtman, and O Wang “The Unreasonable Effectiveness of Deep Features as a Perceptual Metric,” 2018, arXiv:1801. p. 6 Shechtman, and O. Wang, “The
Unreasonable Effectiveness of Deep Features as a Perceptual Metric,”
2018, arXiv:1801.Where the paper says it
--- Page 9 ---
ACKNOWLEDGMENT & CODE
Dr. Alarice Lowe performed pathology interpretations of model outputs and ground truth images Code available at: https://github.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Supervised H&E-to-PIN-4 computational staining framework development enables direct interpretation of predicted PIN-4 marker patterns in context with source H&E architecture by capturing "diagnostically relevant PIN-4 staining patterns, including AMACR/racemase expression and basal-cell-associated staining, while preserving spatial correspondence with the source H&E morphology."
-
The system can perform direct cell-level comparison between the H&E morphology that prompted IHC ordering and the corresponding immunophenotypic signal on an adjacent section by synthesizing PIN-4 staining patterns directly from native H&E image patches.
-
The improved AI system can handle morphologically complex regions, such as high-grade carcinoma and intraductal carcinoma, to capture
diagnostically relevant PIN-4 staining patterns
even when training examples are limited in those areas. -
The model can serve as a potential adjunctive colocalization tool to review predicted PIN-4 marker patterns in spatial context with the source H&E architecture, addressing the limitation that "Conventional adjacent-section IHC cannot guarantee that the glands assessed on PIN-4 are the same glands seen on H&E."
-
The system can be trained and evaluated using a paired, registered dataset comprised of 172 paired WSIs from 93 patients, spanning representation across age, race, and ethnicity groups.
Abstract
Immunohistochemistry (IHC)is frequently used to resolve diagnostically ambiguous prostate cancer biopsy findings on hematoxylin and eosin (H&E)-stained tissue. However, PIN-4 IHC staining is typically performed on adjacent tissue sections, limiting direct spatial comparison between the H&E morphology and the corresponding immunophenotypic signal. A paired, registered H&E/PIN-4 dataset was constructed from routine clinical prostate biopsy whole-slide images (WSIs), and a conditional generative adversarial network (cGAN) was trained to synthesize PIN-4 staining patterns directly from native H&E image patches. The final dataset comprised 172 paired WSIs from 93 patients and 27,298 registered 1024x1024 patch pairs, spanning adenocarcinoma-positive and benign cases with representation across age, race, and ethnicity groups. The model was evaluated on a held-out test set of 1,814 patch pairs from 17 WSIs, achieving a mean peak signal-to-noise ratio (PSNR) of 21.88 dB, structural similarity index measure (SSIM) of 0.667, Pearson correlation coefficient (PCC) of 0.684, and learned perceptual image patch similarity (LPIPS) of 0.417. Qualitative review by a board-certified pathologist showed that generated images captured diagnostically relevant PIN-4 staining patterns, including AMACR/racemase expression and basal-cell-associated staining, while preserving spatial correspondence with the source H&E morphology. Accuracy of synthesis varied across morphologically complex regions, including high-grade carcinoma and intraductal carcinoma. These results support the feasibility of supervised PIN-4 synthesis from routinely acquired brightfield H&E prostate biopsy images. The approach enables direct interpretation of predicted PIN-4 marker patterns in the context of the source prostate H&E architecture, addressing a current spatial limitation of conventional adjacent-section IHC.
Sources
- Tune: A Research Platform for Distributed Model Selection and Training
- The Unreasonable Effectiveness of Deep Features as a Perceptual Metric
- Adam: A Method for Stochastic Optimization
- On the Convergence of Adam and Beyond
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs