Crowdsourcing for pathology

Background

One of the key hurdles preventing translation of advances in machine learning into the medical domain is lack of datasets for algorithm training and validation. Training accurate computational models is dependent on manual delineation (aka. annotation) of tissue regions and other microscopic elements in scanned slides of tumor biopsies. This task that is extremely time consuming and impractical given the time constraints of, and clinical demands on, practicing pathologists.

Aim

Develop algorithmic protocols and computational tools for systematic acquisition for manual delineations of region and cell nuclei from non-pathologists. Explore the accuracy and limitations of annotations from non-experts and develop custom modeling approaches to handle crowd data.

‍