Leveraging Partial Labels for Cervical Lesion Classification via a Multilabel Approach

Early detection and diagnosis of cervical lesions is crucial for mitigating cervical cancer. Because of limited access to adequate annotations, existing studies on cervical lesion classification have primarily focused on generalized classification. This study introduces a multilabel approach to cervical lesion classification through a data-centric framework. We first enhanced data quality via a comprehensive preprocessing pipeline that reduces the dataset to usable cervigrams. This refined dataset is then fed to a Vision Transformer, which performs per-class feature extraction and incorporates a part selection module that emphasizes critical areas within the cervigram, while integrating inter-class information within both early and late stages of the model. These procedures are conducted under the assumption of potential partial labeling. Upon evaluation, our proposed method outperformed existing models, achieving the highest ROC-AUC scores across all lesion grades. These findings suggest that the improved attention mechanisms led to enhanced localization of lesions, enabling focus on the fine details of each lesion. Overall, this study explores the potential of multilabel classification for advancing cervical lesion detection.