Improving Cervical Precancerous Image Segmentation Using Pix2Pix

Abstract

Cervical cancer remains a major health concern worldwide, particularly in developing countries where access to early screening and diagnosis is limited. Visual Inspection with Acetic Acid (VIA) is widely used for cervical cancer screening. However, its interpretation is highly dependent on medical expertise and image quality. Recent advances in deep learning have demonstrated promising performance in medical image segmentation, but their effectiveness is often constrained by limited dataset availability. This study proposes a Pix2Pix-based augmentation approach to improve cervical precancerous image segmentation performance. Cervical precancerous images obtained from RSUP Dr. Mohammad Hoesin Palembang and the International Agency for Research on Cancer (IARC) were annotated into three classes: cervical area, Columnar Area (CA), and lesion area. Two Pix2Pix augmentation configurations were evaluated using image resolutions of 256×256 and 640×640 pixels. The quality of generated synthetic images was assessed using Fréchet Inception Distance (FID), while segmentation performance was evaluated using YOLOv11-seg with Intersection over Union (IoU), Dice Coefficient, Pixel Accuracy, and mean Average Precision (mAP). Experimental results showed that the 640×640 Pix2Pix model achieved a substantially lower FID score of 25.6 compared to 266.0 for the 256×256 configuration, indicating superior image realism. Therefore, only the synthetic images generated by the 640×640 model were incorporated into segmentation training. The augmented dataset improved lesion segmentation performance, achieving increases of 16.4% in mAP50, 12.0% in IoU, and 14.7% in Dice Coefficient compared to the baseline model. These findings demonstrate that high-resolution Pix2Pix augmentation effectively enhances dataset diversity and improves the segmentation performance of YOLOv11-seg for cervical precancerous image analysis.