Z31 - Enabling equitable cervical cancer screening: Automated lesion segmentation across standard and portable colposcopy platforms

Friday, October 10, 2025

10:00 AM - 11:00 AM PDT

Location: Exhibit Hall F, G & H

Presenting Author(s)

Rohan M. Bhansali, BS

Undergraduate Researcher
Duke University, United States

Co-Author(s)

ES

Erica Skerrett

Doctoral Candidate
Duke University, United States

Primary Investigator(s)

Nirmala Ramanujam, PhD

Director, Center for Global Women's Health Technologies, Professor of Biomedical Engineering
Duke University, United States

Introduction: : Cervical cancer disproportionately affects low-resource settings, with 94% of deaths occurring in low- and middle-income countries [World Health Organization, 2024]. While colposcopy remains the gold standard for detecting cervical lesions, traditional colposcopes are expensive and require specialized training. The Pocket Colposcope addresses these barriers as a portable, affordable device enabling point-of-care cervical imaging. However, accurate lesion delineation remains a challenge and is operator-dependent.
Automated segmentation of cervical lesions serves critical clinical purposes, providing biopsy guidance to ensure sampling from the most suspicious areas, supporting classification algorithms by identifying guided landmarks within lesion areas, and enhancing the explainability of diagnostic decisions. This study evaluates SAM-Med2D [Cheng et al., 2023], a foundation model pretrained on 4.6 million medical images and 19.7 million masks, for the automated segmentation of cervical lesions in both Pocket Colposcope and standard colposcopy images. Sam-Med2D employs a vision transformer as its image encoder, enabling efficient feature extraction and segmentation from medical images. Our objective is to develop a robust segmentation pipeline that reduces manual annotation burden while maintaining clinical accuracy. By leveraging a model pretrained on millions of medical images with minimal prompting, we aim to facilitate rapid lesion segmentation during screening procedures. This approach could significantly improve the consistency and efficiency of cervical cancer screening across diverse clinical settings, particularly in resource-limited areas where access to trained colposcopists is scarce.

Materials and

Methods: : We evaluated SAM-Med2D on a dataset of 58 images, comprising Pocket Colposcope (n=38) and standard-of-care (n=20) colposcopy images, which contained zero, one, or multiple lesions with expert ground-truth annotations. The dataset was divided into 44 training and 14 testing images, captured before and after acetic acid application to visualize acetowhite lesions. To handle multiple lesions per image, we modified the preprocessing pipeline to merge individual lesion masks into a single combined annotation. Additionally, we enabled empty mask handling for negative predictions in images without acetowhite areas, crucial for clinical accuracy. The dataset was balanced across lesion characteristics, image quality, and device type.
SAM-Med2D, pre-trained on 4.6 million medical images, was fine-tuned using expert hand-outlined segmentation masks. During training and inference, the model received the original colposcopy image labeled with binary segmentation masks. The vision transformer backbone generates high-dimensional image embeddings, which combine with prompt embeddings (from the prompt encoder) to produce precise segmentation masks through the mask decoder. Training incorporated learnable adapter layers within the transformer blocks to acquire domain-specific knowledge while preserving pretrained features. The 60-epoch training pipeline included data augmentation (resizing/padding) for consistent input dimensions. Performance was evaluated using intersection-over-union (IoU), Dice coefficient, and segmentation loss metrics. The holdout test set included cases with varying lesion counts to assess the model’s ability to handle varying clinical presentations. Experiments were conducted using the best-performing weights on the validation dataset.

Results, Conclusions, and Discussions:: SAM-Med2D achieved promising results, with a test loss of 1.82, an IoU of 0.54, and a Dice coefficient of 0.64 on the held-out data (n=14). 80% of images with one lesion (n=5) had the lesion correctly detected, while 100% of normal images (n=5) were correctly identified as having no lesions. For images containing multiple lesions (n=4), the model detected all lesions in 50% of cases, with smaller peripheral lesions being more likely to be missed. The model reliably captured irregular lesion boundaries, even in challenging conditions, including uneven illumination and specular highlights that are standard in colposcopic imaging. Notably, the model correctly identified images without lesions by accurately distinguishing normal cervical tissue from acetowhite changes.
These findings indicate that foundation models pretrained on medical data can significantly reduce the manual effort required for accurate lesion delineation. The key advantages of SAM-Med2D include its ability to leverage extensive medical pretraining for robust feature recognition, accurate discrimination between normal tissue and acetowhite lesions, and consistent performance across different imaging devices. Notably, the model demonstrated reliable identification of normal cervices without acetowhitening, critical for reducing false positives in screening programs. High image-level sensitivity and specificity, despite moderate pixel-level Dice scores, indicate the model successfully identifies the presence or absence of lesions, with challenges primarily in precise boundary delineation and detection of smaller peripheral lesions. This performance profile aligns well with screening needs, where detecting lesion presence is more critical than perfect segmentation. The minimal prompting requirement and demonstrated performance across both standard and portable colposcopy systems make the approach practical for deployment in diverse clinical settings.
Future work will explore several enhancements, including expanding the dataset size to improve model generalization and performance, automated generation of prompts, clinician-in-the-loop workflows for real-time analysis, image registration pipelines for longitudinal lesion tracking, and multimodal imaging data combining white light and green light images. This work demonstrates the potential of AI-assisted colposcopy to democratize access to quality cervical cancer screening across imaging modalities, reducing the global burden of preventable cervical cancer deaths.

Acknowledgements and/or References (Optional): :