Label Efficient Classification in Liquid Biopsy Data by Self-supervision

Abstract

A cut-off of 5 circulating tumor cells (CTCs) per 7.5 ml of blood in metastatic breast cancer patients is highly predictive of progression-free survival and overall survival. These rare events potentially reflect the disease and heterogeneity of the tumor and therefore need to be isolated and characterized from liquid biopsy samples. Identification of CTCs in peripheral blood samples is only partially automated by the rule-based commercially available Cellsearch® system. However, the system still requires manual post-processing and selection of CTCs from a large number of proposed cell image candidates. We propose a self-supervised (DINO), label-efficient combination of deep learning (DL) and support vector machine (SVM) to reliably identify CTCs. We evaluate the label data efficiency of our method in comparison to supervised DL and show that it consistently outperforms supervised state of the art models in terms of F1 score for different balanced subsets as well as for all available data.