Development of an AI Model for Automated Cervical Cancer Screening Based on Cytology Whole Slide Images

Abstract

Cervical cancer (CC) is the fourth leading cause of cancer-related deaths among women globally, posing a significant health challenge, especially in developing countries with limited healthcare resources. Traditional screening methods like the Pap smear, despite their success in early detection and reduction of CC incidence, face limitations such as labor-intensive analysis and subjective interpretation. Recent technological advancements in artificial intelligence (AI) and computer vision offer promising improvements in CC screening. This study focuses on developing and evaluating a convolutional neural network (CNN) for automated analysis of cervical cytology images from the SIPaKMeD database, which contains 4049 images categorized into five cell types: dyskeratotic, koilocytotic, metaplastic, parabasal, and superficial-intermediate. The proposed CNN architecture includes
convolutional layers, MaxPooling, and Dropout layers to prevent overfitting, optimized using the Adam algorithm. Preprocessing steps such as color normalization, noise reduction, and artifact removal were applied to digitized slide images. The model's performance was assessed using accuracy and recall metrics. Results showed the CNN achieved high accuracy of 92% and sensitivity of 22%, with validation accuracy closely tracking training accuracy, indicating effective learning and generalization. The study demonstrates that Artificial Intelligence (AI) can significantly enhance CC screening accuracy and efficiency, addressing current limitations in traditional methods. Further optimization, including addressing class imbalance and exploring data augmentation techniques, is recommended to enhance the model's predictive capabilities and robustness in diverse clinical settings.