Exploration U-Net Architecture for Cervical Precancerous Lesions Segmentation

Authors

  • Akhiar Wista Arum Universitas Sriwijaya
  • Muhammad Naufal Rachmatullah Universitas Sriwijaya
  • Bambang Tutuko Universitas Sriwijaya
  • Firdaus Universitas Sriwijaya
  • Annisa Darmawahyuni Universitas Sriwijaya
  • Ade Iriani Sapitri Universitas Sriwijaya
  • Anggun Islami Universitas Sriwijaya
  • Dea Agustria Ananda Universitas Sriwijaya

Keywords:

Segmentation, U-Net, Convolutional Neural Network, Cervical Precancerous Lesions

Abstract

The automatic analysis of images for the early detection of cervical cancer relies on the segmentation of cervical precancerous lesions. This paper investigates the incorporation of various CNN-based backbones into a U-Net model for improved segmentation accuracy. A set of twelve backbones was tested, including VGG16, VGG19, ResNet50, ResNext50, EfficientNetB7, InceptionResNetv2, DenseNet201, InceptionV3, MobileNet V2, SE-ResNet50, SE-ResNext50, and SE-Net154. Evaluation metrics were computed using Intersection over Union, pixel accuracy, and Dice coefficient. The findings demonstrate that U-Net with EfficientNetB7 backbone outperforms all other models with an IoU of 73.13%, pixel accuracy of 89.92%, and a Dice coefficient of 77.64%. These results were visually confirmed; segmentation outputs were examined, showing accurate delineation of lesion borders. The dominating performance of EfficientNetB7 was observed to be due to high feature extraction efficiency coupled with powerful spatial information representation. The study is, however, limited by a lack of clinical validation and expert evaluation from trained medical personnel. The results demonstrate the effectiveness of combining the U-Net architecture with advanced CNN backbones towards designing automated systems to analyze medical images.

Downloads

Published

2025-06-01

Issue

Section

Articles