Breast Cancer Tissue Classification from Multiple Annotators using Chained Deep Learning Approaches

Breast cancer is one of the principal causes of cancer death worldwide. The biopsy diagnosis is non-trivial, and specialists often disagree on the final diagnosis. Thus, Computer-aided Diagnosis-(CAD) systems favor the efficiency of this process while reducing the diagnosis time. However, such systems often require large labeled datasets to achieve meaningful performance, which is hard to obtain in medicine. Crowdsourcing approaches deal with this scenario by collecting labels from multiple annotators with varying degrees of expertise. This work explores the application of a multi-annotator for breast cancer tissue classification using a dataset annotated by experts and non-experts. In particular, we tested two loss functions based on a cross-s-entropy function (RCDNN) and a generalized cross-entropy function (GCEDL). Comparative results underscore the challenges posed by the multi-annotator scenario, with the GCECDL model emerging as the most robust, achieving performance levels approaching those of a gold-standard single-annotator model.