Deep Gaussian Processes for Classification With Multiple Noisy Annotators. Application to Breast Cancer Tissue Classification

Machine learning (ML) methods often require large volumes of labeled data to achieve meaningful performance. The expertise necessary for labeling in medical applications like pathology presents a significant challenge developing clinical-grade tools. Crowdsourcing approaches address this by collecting labels from multiple annotators with varying degrees expertise. In recent years, have been adapted learn noisy crowdsourced labels. Among them, Gaussian Processes (GPs) achieved excellent performance due their ability model uncertainty. Deep (DGPs) the limitations GPs using layers enable more complex representations. work, we develop (DGPCR) crowdsourcing problem DGPs first time. DGPCR models (unknown) underlying true labels, and behavior each annotator is modeled confusion matrix among classes. We use end-to-end variational inference estimate both parameters biases. Using annotations 25 pathologists trainees, show that competitive or superior Scalable (SVGPCR) other state-of-the-art deep-learning breast cancer classification. Also, observe obtains better results ( $\text{F}1=81.91$ %) than notation="LaTeX">$\text{F}1=81.57$ deep notation="LaTeX">$\text{F}1=80.88$ curated experts. Finally, an improved estimation annotators’ behavior.