Cervix-aware region-based ensemble graph neural networks for cervical cancer detection from colposcopy images

Abstract

Cervical cancer continues to pose an enormous risk to global health, underscoring the urgent necessity for prompt and accurate detection to improve the health of women affected by this deadly illness. Current advancements on graph neural networks exhibit potential in classifying medical images, yet existing models usually demand massive computational loads, making their implementation unaffordable for clinically resource-limited settings. This research presents an intriguing ensemble learning framework that combines various graph neural network architectures, such as GraphMAE2(large and small variants), GAT, GIN, and GCN, aimed at improving the classification accuracy for cervical cancer identification. Moreover, our approach utilizes the colposcopy images, acquired from the dataset, and transforms them into graph representations enriched with handcrafted medical features and cervix-specific annotations. This provides a detailed and clinically contextualized description of the cervical area. Our ensemble model utilizes an innovative training approach featuring pre-training by masked autoencoding and fine-tuning with cross-validation to arrive at peak performance and generalizability on small datasets. The proposed ensemble model exhibits enhanced performance, attaining a validation accuracy of 95.52% and an F1 score of 94.12%, thereby exceeding the results of individual models. Additionally, we incorporate uncertainty quantification for enhancing the reliability of predictions and it provides insightful feedback on the confidence of predictions. This makes our framework an informative guidance tool for cervical cancer screening, with specific utility for clinically resource-limited settings. This contribution enhances the expanding domain of graph-based processing of medical images and delivers a robust and generalizable solution for healthcare systems.