Visual Explanation of Classification Model Using Prototypical Contrastive Embedding in Cervical Cytology
摘要
In recent years, deep learning algorithms have been developed for classifying atypical cells from liquid-based cytology images. Several studies have emphasized performance, with only a few addressing the interpretation and explanation of classification outcomes. Notably, the visualization of pertinent regions, which is essential for classification models, is pivotal for cytologists to trust and adopt AI-generated results. In this study, we explored the feasibility of visualizing the basis for classification by integrating the concepts of contrastive embedding and prototypical learning into an attention-based model. Consequently, the proposed model demonstrated superior classification performance and provided an informative visual explanation.