# Virtual Digital Histopathology with Explainable Deep Learning for Prostate Cancer Diagnosis

> **NIH NIH R21** · UNIVERSITY OF CALIFORNIA-IRVINE · 2024 · $453,000

## Abstract

Project Summary
Adoption of digital histopathology has increased demands for gigapixel image analysis tools and methods for
clinical applications. Integration of deep learning algorithms for automated tumor segmentation and cancer
diagnosis from Hematoxylin and Eosin (H&E) dye or virtual (stainless) staining of tissue biopsy images have
been reported. There is a significant lack of explanation and performance evaluation (black box phenomenon)
of tumor segmentation and stainless staining deep learning models limiting clinician adoption in oncology
practice. Results from these deep learning systems are known after the biopsy is irreversibly processed by
H&E staining, and do not include confirmatory immunohistochemistry (IHC) biomarker staining diagnosis
performed on adjacent tissue sections. We have previously described deep neural network systems to convert
native nonstained whole slide prostate tissue biopsy images (WSI) to virtual computationally H&E stained
versions validated for tumor segmentation with high precision. The goal of this proposal is developing novel
methods and algorithms for pixel-by-pixel explanations to clinicians and cancer researchers for explainability of
virtual H&E staining augmented with prostate tumor grade segmentation and IHC expression patterns by deep
learning models for digital biopsies. Our previously published and physician authenticated Generative
Adversarial Neural Network (GAN-CS) models trained with 93,199 image pairs of prostate biopsy images for
virtual H&E staining and prostate tumor segmentations, and a publicly available database of clinically validated
IHC analysis of prostate biopsy images will be used to generate the explainability software. A histological map
that visualizes and interprets correspondence between neural representations in GAN-CS model with virtually
stained prostate Gleason grade tumors in WSI at a pixel level will be generated. A separate explainable GAN-
CSS model for segmentation and Gleason grading of tumors using clinician annotations from H&E and
conjunctive IHC image biomarker labels will also be generated. Researchers can upload WSI into GAN-CS/S
software to perform computational H&E staining multiplexed with morphological tumor segmentations and IHC
expression with pixel-by-pixel visualization and explanation to characterize cellular phenotypes for cancer
research and accelerate histopathology diagnoses. The open source software toolkit developed from this
research can be generalized to majority of deep learning model architectures using disease labels from widely
available non-stained, chemical, or virtual H&E and IHC stained images from different cancer types.

## Key facts

- **NIH application ID:** 10990125
- **Project number:** 1R21CA294195-01
- **Recipient organization:** UNIVERSITY OF CALIFORNIA-IRVINE
- **Principal Investigator:** Pratik Shah
- **Activity code:** R21 (R01, R21, SBIR, etc.)
- **Funding institute:** NIH
- **Fiscal year:** 2024
- **Award amount:** $453,000
- **Award type:** 1
- **Project period:** 2024-09-01 → 2026-08-31

## Primary source

NIH RePORTER: https://reporter.nih.gov/project-details/10990125

## Citation

> US National Institutes of Health, RePORTER application 10990125, Virtual Digital Histopathology with Explainable Deep Learning for Prostate Cancer Diagnosis (1R21CA294195-01). Retrieved via AI Analytics 2026-05-24 from https://api.ai-analytics.org/grant/nih/10990125. Licensed CC0.

---

*[NIH grants dataset](/datasets/nih-grants) · CC0 1.0*