Developing a Deep Learning-Based Model for Cervical Cancer Screening Using Pap Smear Images

Abstract

Cervical cancer remains a major health concern, especially in regions with limited screening resources. While deep learning offers promising solutions for Pap smear image analysis, model performance is constrained by the scarcity of high-quality pixel-level annotations. This thesis proposes an iterative semi-supervised segmentation framework built on a standard U-Net, combining confidence-aware pseudo-label refinement with a Smooth Threshold Evolution (STE) strategy. Experiments on the SIPaKMeD dataset demonstrate that the proposed approach improves mean Dice from 0.6370 to 0.6644 on an independent clean test set, while Cytoplasm Recall increases from 0.4529 to 0.6186. These results are achieved without architectural modifications or large-scale pretraining, proving that iterative pseudo-label refinement is a practical and accessible strategy for medical image segmentation under realistic annotation constraints.