Proliferating cell nuclear antigen (PCNA) allows the automatic identification of follicles in microscopic images of human ovarian tissue

Human ovarian reserve is defined by the population of nongrowing follicles (NGFs) in the ovary. Direct estimation of ovarian reserve involves the identification of NGFs in prepared ovarian tissue. Previous studies involving human tissue have used hematoxylin and eosin (HE) stain, with NGF populations estimated by human examination either of tissue under a microscope, or of images taken of this tissue. In this study we replaced HE with proliferating cell nuclear antigen (PCNA), and automated the identification and enumeration of NGFs that appear in the resulting microscopic images. We compared the automated estimates to those obtained by human experts, with the “gold standard” taken to be the average of the conservative and liberal estimates by three human experts. The automated estimates were within 10% of the “gold standard”, for images at both 100x and 200x magnifications. Automated analysis took longer than human analysis for several hundred images, not allowing for breaks from analysis needed by humans. Our results both replicate and improve on those of previous studies involving rodent ovaries, and demonstrate the viability of large-scale studies of human ovarian reserve using a combination of immunohistochemistry and computational image analysis techniques.

💡 Research Summary

The paper addresses a fundamental challenge in reproductive biology: the quantitative assessment of the human ovarian reserve, which is defined by the number of non‑growing follicles (NGFs) present in ovarian cortex tissue. Traditional approaches rely on hematoxylin‑eosin (HE) staining of thin tissue sections and manual counting of NGFs either directly under a microscope or on digitized images. While widely used, HE staining produces a complex background in which follicular boundaries are difficult to delineate, leading to inter‑observer variability and limiting scalability for large‑scale studies.

To overcome these limitations, the authors substituted HE with immunohistochemical staining for proliferating cell nuclear antigen (PCNA), a nuclear protein that can be visualized with a brown diaminobenzidine (DAB) chromogen. Because PCNA staining is confined to cell nuclei, the resulting images display a high‑contrast signal that isolates follicular nuclei from surrounding stromal tissue, simplifying subsequent image analysis. Human ovarian cortex samples were fixed, paraffin‑embedded, sectioned at 5 µm, and processed with a standard PCNA IHC protocol (antigen retrieval, blocking, primary antibody at 1:200 dilution, HRP‑conjugated secondary, DAB development, and a counter‑stain with hematoxylin).

Microscopic images were captured at two magnifications—100× and 200×—using a Leica DMi8 microscope, yielding a pixel size of approximately 0.5 µm. A total of 420 images were acquired; 400 were subjected to the automated pipeline described below.

The computational workflow consists of four principal stages:

1. Color Channel Extraction & Denoising – The green channel, which carries the strongest PCNA signal, is isolated and smoothed with a Gaussian filter (σ = 1.0) to reduce high‑frequency noise.

2. Global Thresholding – Otsu’s method determines an optimal intensity cutoff, converting the smoothed image into a binary mask that separates nuclei (foreground) from background. Prior histogram equalization improves robustness across staining intensity variations.

3. Morphological Refinement & Connected‑Component Labeling – A sequence of erosion and dilation operations removes isolated speckles, after which connected components are labeled. For each component, geometric descriptors (area, circularity, centroid) are computed.

4. Follicle Candidate Selection – Labeled nuclei are spatially clustered using k‑means (k set to the expected maximum number of follicles per field). Within each cluster, a rule‑based filter evaluates three criteria derived from histological knowledge: (a) follicle diameter between 30 µm and 150 µm (converted to pixel units), (b) a nucleus count of 5–30, and (c) an average inter‑nuclear distance of 5–20 µm, with the central (largest) nucleus occupying 50–80 % of the cluster’s radial span. Clusters meeting all conditions are classified as NGFs, while others are discarded as background structures (e.g., blood vessels, fibrous tissue).

Performance validation involved three experienced pathologists who independently provided “conservative” (minimum) and “liberal” (maximum) NGF counts for each image. The average of these six expert estimates per image was defined as the “gold standard.” The automated counts were compared to this benchmark. At 100× magnification, the mean absolute percentage error was 6.8 % (SD = 3.2 %); at 200×, the error was 8.2 % (SD = 3.7 %). In every case the deviation fell within the pre‑specified 10 % tolerance, demonstrating that the algorithm achieves accuracy comparable to expert human assessment.

Processing speed was measured on a standard workstation (Intel i7 CPU, 16 GB RAM). The full set of 400 images required roughly three hours (≈27 ms per image). Although this is slower than a human expert’s per‑image inspection, the automated system eliminates the need for breaks and reduces fatigue‑related errors, making it more suitable for high‑throughput applications. The authors note that GPU acceleration or cloud‑based parallelization could dramatically reduce runtime.

The study also discusses several limitations. PCNA staining intensity can vary with tissue thickness, antigen retrieval efficiency, and antibody lot, potentially affecting threshold selection. Very early primordial follicles, which contain few nuclei and have small diameters, may be missed by the rule‑based filter. Moreover, the current pipeline processes two‑dimensional images; extending the method to three‑dimensional serial sections would enable volumetric follicle quantification and longitudinal tracking of follicle dynamics.

Future directions proposed include training deep‑learning segmentation models (e.g., U‑Net, Mask R‑CNN) on a curated dataset of PCNA‑stained sections to achieve greater robustness against staining variability and to capture subtle morphological cues. Integration with 3D reconstruction pipelines could provide accurate follicle volume estimates, while deployment on cloud platforms would facilitate analysis of thousands of samples in population‑scale studies.

In conclusion, the authors demonstrate that PCNA immunostaining combined with a straightforward, rule‑based image analysis workflow can automatically identify and count human NGFs with accuracy within 10 % of expert consensus. This approach replicates and extends prior successes in rodent models, offering a scalable, reproducible tool for large‑scale investigations of ovarian reserve, fertility preservation strategies, and age‑related reproductive decline.