Automated Cervical Cancer Screening Using Single-Cell Segmentation and Deep Learning: Enhanced Performance with Liquid-Based Cytology

Cervical cancer (CC) remains a significant health issue, especially in low- and middle-income countries (LMICs). While Pap smears are the standard screening method, they have limitations, like low sensitivity and subjective interpretation. Liquid-based cytology (LBC) offers improvements but still relies on manual analysis. This study explored the potential of deep learning (DL) for automated cervical cell classification using both Pap smears and LBC samples. A novel image segmentation algorithm was employed to extract single-cell patches for training a ResNet-50 model. The model trained on LBC images achieved remarkably high sensitivity (0.981), specificity (0.979), and accuracy (0.980), outperforming previous CNN models. However, the Pap smear dataset model achieved significantly lower performance (0.688 sensitivity, 0.762 specificity, 0.8735 accuracy). This suggests that noisy and poor cell definition in Pap smears pose challenges for automated classification, whereas LBC provides better classifiable cells patches. These findings demonstrate the potential of AI-powered cervical cell classification for improving CC screening, particularly with LBC. The high accuracy and efficiency of DL models combined with effective segmentation can contribute to earlier detection and more timely intervention. Future research should focus on implementing explainable AI models to increase clinician trust and facilitate the adoption of AI-assisted CC screening in LMICs.