Unreliable Annotation Detection with Similarity Graph in Cervical Cytology

We propose a novel method for detecting unreliable annotations, which are often referred to as low-quality labels obscured within large datasets of cervical cytology images. Our approach entails building a similarity graph from the dataset and then identifying a dominant label within each graph clique. This strategy of focusing on dominant labels is specifically designed to enhance subsequent label propagation performance. A key advantage of our method is its ability to reliably identify low-quality labels, even those dispersed across multiple categories, while achieving stable accuracy without relying on any supervised information. We experimentally validated our method, demonstrating its successful detection of pseudo-label errors that were intentionally and randomly injected into the original datasets. Furthermore, our observations confirm that identifying dominant labels within graph cliques improves the error detection rate. The primary contribution of our work lies in proposing the first method for selecting initial labels using similarity graphs and label propagation, and in demonstrating its ability to achieve high accuracy despite its inherent simplicity. Consequently, our method can be readily integrated into practical annotation support systems, offering annotators enhanced visibility and facilitating a human-in-the-loop annotation process.