Assessing the diagnostic accuracy of artificial intelligence in detecting cervical pre-cancer from pap smear images

The global burden of cervical cancer, with a notable prevalence in regions like Tanzania, highlights the critical need for timely and accurate diagnosis. The scarcity of pathologists in such areas underscores the importance of developing automated tools for the cytopathological analysis of cervical cancer images to improve diagnostic efficiency and accuracy. This study investigated the performance of advanced artificial intelligence (AI) algorithms for screening cervical cancer using Pap smear cytological slides from the Centre for Recognition and Inspection of Cells CRIC dataset. Deep learning models were trained on images of both cervical cancer and normal cervix cells, with the evaluation of model performance focusing on specificity, sensitivity, and accuracy. Among the evaluated convolutional neural network (CNN) architectures--EfficientNetB7, MobileNet, ResNet50, ResNet152, and InceptionNet-V3--EfficientNetB7 emerged as the top performer, demonstrating impressive accuracy, sensitivity and specificity metrics (accuracy: 0.95, sensitivity: 0.84, specificity: 0.97). In contrast, InceptionNet-V3 showed the lowest performance across similar metrics (accuracy: 0.78, sensitivity: 0.35, specificity: 0.87). The study also highlighted the challenges in distinguishing between specific cell classes, particularly between ASC-US and LSIL, due to the subtle nuances of cytomorphological criteria. The findings suggest that while AI can significantly aid in cervical cancer screening, the complexity of cell image classification demands further exploration, possibly incorporating whole smear analysis or additional contextual information to improve accuracy. Despite the observed classification challenges, such as between ASC-US and LSIL classes, the potential for AI in supporting clinical decision-making in cervical cancer management is evident.