Automated Cervical Os Segmentation for Camera-Guided, Speculum-Free Screening

Aoife McDonald-Bowyer, Anjana Wijekoon, Ryan Laurance Love, Katie Allan, Scott Colvin, Aleksandra Gentry-Maharaj, Adeola Olaitan, Danail Stoyanov, Agostino Stilli, Sophia BanoarXiv 2025 · arXiv

TL;DR

To make cervical cancer screening more accessible and comfortable, researchers are developing speculum-free, camera-guided sampling devices. This paper evaluates five deep learning models for real-time localization of the cervical os (the opening of the cervix), finding that a vision transformer pre-trained on surgical video (EndoViT/DPT) delivers the most reliable performance at 21.5 frames per second.

What Problem Does This Paper Address?

Cervical cancer is highly preventable through early detection, yet screening rates face major hurdles globally. In low- and middle-income countries (LMICs), a shortage of trained clinicians limits access to traditional cytology and colposcopy. In high-income countries, patient uptake is declining, frequently due to the discomfort and anxiety associated with speculum-based examinations.

While primary HPV self-sampling allows patients to collect samples blindly, a positive result still requires an in-clinic follow-up. To streamline this pathway, researchers are developing speculum-free devices that combine imaging and cell collection in a single step. However, operating these devices without a speculum means the cervix isn't mechanically exposed. To allow non-experts to safely navigate the device and align the collection brush, automated, real-time visual guidance is required to locate the cervical os. This paper addresses the challenge of accurately segmenting the cervical os from transvaginal endoscopic images.

Methods

The authors benchmarked five state-of-the-art encoder-decoder architectures to find the best balance between segmentation accuracy and inference speed:

  • EndoViT/DPT: A vision transformer pre-trained on a large collection of endoscopic surgical videos and fine-tuned on laparoscopic data.
  • DeepLabV3: A popular CNN architecture utilizing atrous (dilated) convolutions and Atrous Spatial Pyramid Pooling (ASPP) to capture multi-scale context.
  • PSPNet: A CNN that uses a pyramid pooling module to gather global context embedded with a ResNet backbone.
  • YOLOv8 & YOLOv11: Highly optimized, real-time models trained initially on the COCO dataset, with YOLOv11 serving as an experimental transformer-based variant.

Performance was evaluated using ten-fold cross-validation across standard segmentation metrics: Intersection over Union (IoU), DICE score, Detection Rate (DR), Centroid Distance (CenD), and Minimum Distance (MinD).

Datasets

The models were trained and validated using two distinct datasets:

  1. IARC Cervical Image Dataset: A public dataset from which the authors extracted 913 frames ($800 \times 600$ pixels) spanning 200 clinical cases. Three gynecologists provided ground-truth, pixel-wise annotations of the cervical os.
  2. Custom Phantom Dataset: For external validation, the authors fabricated a silicone cervico-vaginal phantom based on human anatomical dimensions. They recorded a 70-second video ($1280 \times 720$ resolution at 30 fps) using a 2 mm USB endoscope inside a prototype speculum-free device. From this, 70 frames (sampled at 1 fps) were used to test how well the models generalized to unseen, hardware-specific data.

Key Insights and Ideas

  • Domain-Specific Pre-training Wins: The vision transformer (EndoViT/DPT) outperformed the other models in detection consistency. The authors attribute this to its pre-training on surgical videos, which helps the model better understand the unique textures, specular reflections, and lighting conditions of endoscopic imagery compared to models trained on standard datasets like COCO.
  • Real-Time Feasibility: EndoViT/DPT achieved an inference speed of 46.5 ms per frame (~21.5 FPS). This demonstrates that transformer-based architectures are fast enough to provide live visual feedback during a procedure.
  • Robustness to Occlusion: External validation on the phantom data showed that the transformer model could successfully localize the os even when it was off-center or partially occluded—a critical requirement for a device intended for non-expert self-sampling.

Key Results

  • EndoViT/DPT achieved the highest DICE score ($0.50 \pm 0.31$) and Detection Rate ($0.87 \pm 0.33$), indicating the most consistent overlap with the expert annotations.
  • DeepLabV3 secured the highest IoU ($0.40 \pm 0.28$), though its overall DICE and detection rates were slightly lower than the vision transformer.
  • The YOLO-based models generally demonstrated weaker performance, particularly in detection sensitivity.
Quantitative performance comparison across the five evaluated models.

Qualitative analysis highlighted that while most models succeeded in clear, centered cases, they struggled with edge cases involving ambiguous anatomy or lighting artifacts.

Qualitative segmentation examples across different cases, showing success (Cases 1–3) and failure modes (Cases 4–6).

During phantom testing, EndoViT/DPT successfully identified the external os under variable positioning and partial occlusion conditions.

Fig. 2. Qualitative segmentation results using EndoViT/DPT on phantom data recorded with the self-sampling device under development. Ground truth (in Green), model predictions (in Blue) and performance metrics ([IoU,DICE] as percentages) are overlaid to indicate model performance.

Limitations and Future Work

The authors noted a minor evaluation bias: the Centroid Distance (CenD) and Minimum Distance (MinD) metrics were only calculated when both the ground truth and the model's predictions were present, meaning total detection failures did not penalize these specific distance scores.

Additionally, while the detection rate was high, the absolute DICE scores (~0.50) indicate that there is still significant room to refine boundary accuracy. Future work will focus on task-specific hyperparameter tuning and embedding these segmentation capabilities directly into the hardware of the prototype speculum-free imaging and sampling system currently under development.

Summary updated

Abstract

Cervical cancer is highly preventable, yet persistent barriers to screening limit progress toward elimination goals. Speculum-free devices that integrate imaging and sampling could improve access, particularly in low-resource settings, but require reliable visual guidance. This study evaluates deep learning methods for real-time segmentation of the cervical os in transvaginal endoscopic images. Five encoder-decoder architectures were compared using 913 frames from 200 cases in the IARC Cervical Image Dataset, annotated by gynaecologists. Performance was assessed using IoU, DICE, detection rate, and distance metrics with ten-fold cross-validation. EndoViT/DPT, a vision transformer pre-trained on surgical video, achieved the highest DICE (0.50 \pm 0.31) and detection rate (0.87 \pm 0.33), outperforming CNN-based approaches. External validation with phantom data demonstrated robust segmentation under variable conditions at 21.5 FPS, supporting real-time feasibility. These results establish a foundation for integrating automated os recognition into speculum-free cervical screening devices to support non-expert use in both high- and low-resource contexts.
BibTeX
@misc{mcdonaldbowyer2025automated,
  author = {Aoife McDonald-Bowyer and Anjana Wijekoon and Ryan Laurance Love and Katie Allan and Scott Colvin and Aleksandra Gentry-Maharaj and Adeola Olaitan and Danail Stoyanov and Agostino Stilli and Sophia Bano},
  title = {{Automated Cervical Os Segmentation for Camera-Guided, Speculum-Free Screening}},
  year = {2025},
  eprint = {2509.10593},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.10593},
}

Loading related content...

Figure