Artificial Intelligence-Assisted Colposcopy: Deep Learning Multi-Class Segmentation of Anatomical Structures and Pathological Findings for Cervical Cancer Screening.

Background

Accurate colposcopy assessment is essential for detection of cervical precancerous lesions. However, the diagnostic performance depends heavily on the examiner's experience and subjective interpretation. Recent advances in artificial intelligence offer new opportunities for automated image analysis. In particular, deep learning models have shown promise for colposcopy image analysis intended for precancer screening. However, their limited interpretability restricts their translational value in clinical gynecology. This study leverages a custom dataset of expert-annotated digital colposcopic images to quantify the diagnostic strengths of these architectures.

Methods

A comparative analysis was provided of the well-established convolutional neural network YOLOv11 and the modern transformer-based RF-DETR architecture, both of which were trained for the segmentation of 10 distinct classes. Target objects included anatomical structures, medical instruments and colposcopic findings.

Results

Our results demonstrate that the YOLO architecture provides better performance for anatomical structures, whereas the RF-DERT reaches higher scores for colposcopy findings. These findings demonstrate transformer architecture superiority in more nuanced segmentation of clinically relevant findings, providing an interesting framework for computer-aided decision support systems in colposcopy.

Conclusions

Both YOLOv11 and RF-DETR enable effective segmentation of colposcopic images, with performance dependent on class size and characteristics. Best results were achieved for large anatomical structures, while small and underrepresented findings remained challenging due to class imbalance. YOLO offers greater stability and efficiency, whereas RF-DETR performs better on more complex cases. Limitations include data imbalance, variable image quality, and annotation inconsistencies; future work should address these issues to improve generalization.