Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2025 May 23;9(1):151.
doi: 10.1038/s41698-025-00934-5.

Mitigating bias in prostate cancer diagnosis using synthetic data for improved AI driven Gleason grading

Derek J Van Booven  1 , Cheng-Bang Chen  2 , Oleksandr N Kryvenko  3   4 , Sanoj Punnen  4   5 , Victor Sandoval  6 , Sheetal Malpani  3 , Ahmed Noman  7 , Farhan Ismael  8 , Yujie Wang  2 , Rehana Qureshi  3 , Joshua M Hare  9   10   11 , Himanshu Arora  12   13   14   15
Affiliations

Mitigating bias in prostate cancer diagnosis using synthetic data for improved AI driven Gleason grading

Derek J Van Booven et al. NPJ Precis Oncol. .

Abstract

Prostate cancer (PCa) is a leading cause of cancer-related mortality in men, with Gleason grading critical for prognosis and treatment decisions. Machine learning (ML) models offer potential for automated grading but are limited by dataset biases, staining variability, and data scarcity, reducing their generalizability. This study employs generative adversarial networks (GANs) to generate high-quality synthetic histopathological images to address these challenges. A conditional GAN (dcGAN) was developed and validated using expert pathologist review and Spatial Heterogeneous Recurrence Quantification Analysis (SHRQA), achieving 80% diagnostic quality approval. A convolutional neural network (EfficientNet) was trained on original and synthetic images and validated across TCGA, PANDA Challenge, and MAST trial datasets. Integrating synthetic images improved classification accuracy for Gleason 3 (26%, p = 0.0010), Gleason 4 (15%, p = 0.0274), and Gleason 5 (32%, p < 0.0001), with sensitivity and specificity reaching 81% and 92%, respectively. This study demonstrates that synthetic data significantly enhances ML-based Gleason grading accuracy and improves reproducibility, providing a scalable AI-driven solution for precision oncology.

PubMed Disclaimer

Conflict of interest statement

Competing interests: Disclosure of Patent Information: The authors wish to inform that the technology presented in this study is part of a provisional patent application that has been filed with the United States Patent and Trademark Office (USPTO). The application has been assigned Serial No. 63/598,207 and was filed on November 13, 2023. The patent application is currently pending. Some of the authors of this paper are listed as inventors in the patent application. This patent filing may constitute a potential conflict of interest, and this statement serves to disclose this relationship in the interest of full transparency. Joshua M. Hare reports having a patent for cardiac cell-based therapy and holds equity in Vestion Inc., and maintains a professional relationship with Vestion Inc. as a consultant and member of the Board of Directors and Scientific Advisory Board. Vestion Inc. did not play a role in the design, conduct, or funding of the study. Dr. Joshua Hare is the Chief Scientific Officer, a compensated consultant, and a board member for Longeveron Inc. and holds equity in Longeveron. Dr. Hare is also the co-inventor of intellectual property licensed to Longeveron. Longeveron did not play a role in the design, conduct, or funding of the study. The University of Miami is an equity owner in Longeveron Inc., which has licensed intellectual property from the University of Miami. Ethics statement: This study was conducted in accordance with the principles outlined in the Declaration of Helsinki. Ethical approval was obtained from the Institutional Review Board (IRB) of the University of Miami Miller School of Medicine, Miami, FL (IRB Protocol Number: 20140372). The MAST Trial was registered on ClinicalTrials.gov (Identifier: NCT02242773). Informed consent was obtained from all participants prior to their inclusion in the study. Additionally, external datasets, including those from The Cancer Genome Atlas (TCGA), Radboud University Medical Center, and Karolinska Institute (PANDA challenge), were used in compliance with their respective data use agreements. All data were anonymized to ensure participant confidentiality and privacy.

Figures

Fig. 1
Fig. 1. Overview of synthetic image generation workflow from prostate cancer histology.
A Illustration showing pipeline used in generating synthetic images from prostate cancer digital histology. Images were preprocessed by PyHist and HistoQC. Those that passed QC were then given to pathologist for scoring, and then cut into small patches for modeling. B Original and Synthetic images were generated for each primary Gleason pattern 3, 4, and 5, respectively.
Fig. 2
Fig. 2. GAN training and synthetic biopsy image generation pipeline.
A Workflow for needle biopsy images that were used in developing the training database to be used in the GAN. Images were normalized, then fed into the GAN, and then assessed for quality. B Example original and synthetic histology images generated for prostate cancer needle biopsies.
Fig. 3
Fig. 3. Comparison of spatial recurrence features in real and synthetic radical prostatectomy images.
A The distributions of spatial recurrence properties (in the first 6 Principal Components (PCs), which contain 82% of data variability) underlying different Gleason scores for both real and synthetic patches on Radical Prostatectomy. Note that the purple lines indicate the mean values of each feature, and the gray area shows the 95% confidence interval. Our results indicate that while the distributions of spatial properties are closely aligned between real and synthetic images under the same Gleason Score, they markedly differ when comparing different Gleason Scores. B The comparison of spatial recurrence properties between real and synthetic on the first six PCs (contain 82% of data variability). The distributions of this four PCs are similar between real and synthetic.
Fig. 4
Fig. 4. Validation of spatial recurrence consistency in synthetic needle biopsy images.
A The distributions of spatial recurrence properties (in the first 16 Principal Components (PCs), which contain 80% of data variability) underlying different Gleason scores for both real and synthetic patches on Needle Biopsy. Note that the purple lines indicate the mean values of each feature, and the gray area shows the 95% confidence interval. Our results indicate that while the distributions of spatial properties are closely aligned between real and synthetic images under the same Gleason Score, they markedly differ when comparing different Gleason Scores. B The comparison of spatial recurrence properties between real and synthetic on the first eight PCs (contain 70% of data variability). The distributions of this four PCs are similar between real and synthetic.
Fig. 5
Fig. 5. SHRQA-derived granular features distinguishing Gleason patterns in synthetic images.
Shows the distributions of granular features associated with (A) Gleason pattern 3, (B) Gleason pattern 4, and (C) Gleason pattern 5 as identified by the SHRQA quantification and verified by the pathologists.
Fig. 6
Fig. 6
Showing cumulative improvement in accuracy through ROC curves between synthetic+original against the original dataset of RP (left) and needle biopsies (right). p < 0.05 in both cases.
See this image and copyright information in PMC

References

    1. Brawley, O. W. Prostate cancer epidemiology in the United States. World J. Urol.30, 195–200 (2012). - PubMed
    1. Badalament, R. A. & Drago, J. R. Prostate cancer. Dis. Mon.37, 199–268 (1991). - PubMed
    1. Carthon, B., Sibold, H. C., Blee, S. & D Pentz, R. Prostate Cancer: Community Education and Disparities in Diagnosis and Treatment. Oncologist26, 537–548 (2021). - PMC - PubMed
    1. Cook, E. D. & Nelson, A. C. Prostate cancer screening. Curr. Oncol. Rep.13, 57–62 (2011). - PubMed
    1. Litwin, M. S. & Tan, H. J. The Diagnosis and Treatment of Prostate Cancer: A Review. Jama317, 2532–2542 (2017). - PubMed

LinkOut - more resources