Unsupervised Learning with Label Commonization for Uncovering Low-Quality Labels

We propose a label commonization technique with transformed label propagation to uncover low-quality labels which are assigned by a human annotation task in the medical dataset. Our challenge lies in identifying specific labeled data hidden from all data with multiple categories without any supervised information. Our pioneering work here focuses on unsupervised graph-based machine learning as a basis and applies it to search low-quality labels from enormous medical image datasets so as not to be used in machine learning model training. We evaluated our method in three types of cervical cancer image datasets and found its effectiveness in finding low-quality labels with quite high accuracy. Ablation studies highlight the importance of parameters in graph-related processes.