Graph-Based Similarity Search for Robust Annotation in Cervical Cytology

We propose a novel annotation support method for annotators to label a cervical cytology image dataset according to its cancer progress stage with a graph-based similarity search. While annotation support tools for multiple image classification especially in medical fields are an imminent requirement for medical diagnosis, there are few studies or implementations of these types of tools due to their substantial difficulties. Our work provides annotators with a practical reference by retrieving similar images with their labels in a Top-k search, which guides annotators to assign a label to a target image. A unique point of our study lies in building a graph with image similarity and providing a similarity search of images with assigned labels by local structural similarity in the graph. The key advantage of our method is that it allows annotators to perform careful annotation by providing the most probable label for the target image together with other probable candidate labels for comparison, which offers annotators high visibility with explainable search results. Despite its simplicity, our method demonstrates quite good accuracy in suggesting correct labels compared to the naïve application of image similarity with the Vision Transformer.