Pathology image analysis with topological data analysis and foundation models

抄録

Pathology images are usually classified by morphology, but this approach does not fully capture the ambiguous and continuous nature of cancer tissues. To better represent this, we need a distribution that interpolates classifications. However, it remains unclear which image quantification is best suited for this purpose and what evaluation criteria we should use in the absence of human-generated ground truth. In this study, we set a pipeline for the objective evaluation of distributions and test frameworks based on foundation models and topological data analysis. Our results suggest the effectiveness of topological information in enhancing performance when combined with foundation models.