Self-Supervised Learning and Active Learning Pipeline for Cervical Cytology Clustering
Gerardo Escoto & Matthew Flowerhill
Lay Summary:
We developed a computer-based tool that organizes cervical screening images by grouping together cells that look similar, helping experts focus their attention on the cells that may be most important to review. Our goal is to make cervical cancer screening faster and more efficient, especially in places like Honduras where access to trained specialists and screening resources can be limited.
Abstract:
Cervical cancer remains a major public health challenge, particularly in resource-limited settings where limited access to expert cytopathologists can delay diagnosis and reduce screening efficiency. We developed an end-to-end computational pipeline for cervical cytology whole-slide images that integrates YOLO-based cell detection, quality-controlled preprocessing, self-supervised teacher–student Vision Transformer learning (CytoFM), graph-based clustering, and active learning prioritization. Individual cell crops are transformed into high-dimensional feature embeddings that capture morphology and enable unsupervised graph-based community detection of morphologically similar cells, while an active learning framework prioritizes informative cells for expert review. Together, this pipeline provides a scalable workflow for AI-assisted digital pathology by reducing manual review effort and improving the efficiency of expert analysis.
Q&A:
Bios: Gerardo Escoto,Matthew Flowerhill
Program Track: Skills Development
GitHub Username:
GerardoE3737 -Gerardo Escoto
mattjunior039 -Matthew Flowerhill
What was your favorite seminar? Why?
My favorite seminar was the “Cytopathology Fundamentals and Thyroid Cytology AI Research in Progress”. I thought that this seminar was interesting because it kind of correlated to the research that Matthew and I were doing. We worked with cervical cytology images so hearing about thyroid cytology was a fun experience for me. It gave me a different perspective on the whole study of cytology and made me open to looking into other specific fields that deal with cytololgy. -Gerardo Escoto
My favorite seminar was the one on computational pathology as it provided a very basic, intuitive understanding of neural networks and helped me understand the shape of the field. -Matthew Flowerhill
If you were to summarize your summer internship experience in one sentence, what would it be?
My summer internship was very refreshing and informative -Gerardo Escoto
My summer internship experience can be described as a very extensive learning experience, as I learned valuable skills from amazing mentors that helped me with my work. -Matthew Flowerhill
Blog Post
Gerardo Escoto
Matthew Flowerhill
25 August 2026
Self-Supervised Learning and Active Learning Pipeline for Cervical Cytology Clustering
Introduction
Cervical cancer remains a major global health challenge, with the greatest burden occurring in low- and middle-income countries where access to vaccination, screening, and treatment is often limited. These disparities highlight the need for more accessible and efficient approaches to cervical cancer screening. Our project explores how artificial intelligence and digital cytology can help address this challenge by organizing cervical cells based on their visual characteristics and prioritizing potentially informative cells for expert review.
Background
Honduras faces significant challenges in cervical cancer prevention, screening, and treatment. Cervical cancer is the leading cause of cancer-related death among Honduran women ages 15–44, with approximately 858 new cases and 518 deaths each year. Limited access to trained pathologists and the centralization of healthcare services can create major bottlenecks in diagnosis and treatment, with reported treatment delays reaching approximately 232 days. Poverty and geographic isolation can further restrict access to routine screening, particularly for women living in rural communities.
Traditional cervical cancer screening methods can also be difficult to implement effectively in low-resource settings. Pap smears require trained personnel and laboratory infrastructure, while visual inspection methods can produce high rates of false-positive results. HPV testing can identify individuals at increased risk, but positive results still require additional evaluation and triage. These limitations create an opportunity for digital cytology and computational tools to help organize large amounts of cervical cell data and support a more efficient expert review process.
Methodology
To address these challenges, we developed a computational pipeline using digital cervical cytology whole-slide images from La Liga Medical Center in Honduras. The pipeline begins by using a YOLO-based model to detect and segment individual cells from each image. A quality-control step then evaluates the detected cells using model confidence and morphological characteristics, such as cell shape and size, to help filter out artifacts and unreliable detections.
The remaining cells are analyzed using CytoFM, a Vision Transformer trained through self-supervised learning. This approach allows the model to learn meaningful features of cellular morphology without requiring every cell to be manually labeled. These learned representations are then used to group morphologically similar cells through k-nearest neighbor graph construction and Leiden clustering. Finally, an active-learning system scores cells based on uncertainty, rarity, and diversity, creating a prioritized review queue designed to direct expert attention toward potentially informative cells.
Analysis/Results
{width=”6.5in” height=”5.791666666666667in”}
The pipeline successfully organized cervical cell images into distinct groups based on their learned morphological features. When the cell representations were visualized using UMAP, clear groupings appeared across the dataset, suggesting that the model was able to capture meaningful differences in cell morphology. Silhouette scores were also used to evaluate the quality of these clusters by measuring how closely cells within the same cluster were grouped compared with cells in other clusters.
The active-learning component further improved the efficiency of the review process by changing which cells were prioritized for expert evaluation. Rather than selecting large numbers of cells from common and repetitive clusters, the system reduced their representation and placed greater emphasis on rare, uncertain, or underrepresented groups. This created a more balanced review queue designed to maximize the amount of useful information presented to an expert.
Conclusion
Overall, this project demonstrates the potential of combining self-supervised learning, cell clustering, and active learning to support large-scale cervical cytology analysis. The developed pipeline can process cervical cell images, learn important morphological features without relying on extensive manual annotations, and organize cells into distinct groups. By prioritizing rare, uncertain, and diverse cells for expert review, the approach may help reduce repetitive manual review and allow specialists to focus their attention on potentially more informative cells.
Future work will focus on further fine-tuning the Vision Transformer to improve its ability to represent cervical cell morphology and improving the YOLO model's accuracy in distinguishing true cells from imaging artifacts. The dataset will also be expanded to include cervical cytology images from additional hospitals and clinical settings. Increasing the diversity of the data will be important for evaluating how well the pipeline generalizes across differences in staining, imaging conditions, and patient populations. With continued development and validation, this work provides a foundation for more scalable and efficient digital cytology workflows, particularly in settings where access to trained specialists is limited.