Active Data Enrichment by Learning What to Annotate in Digital Pathology

Our work aims to link pathology with radiology with the goal to improve the early detection of lung cancer. Rather than utilising a set of predefined radiomics features, we propose to learn a new set of features from histology. Generating a comprehensive lung histology report is the first vital step toward this goal. Deep learning has revolutionised the computational assessment of digital pathology images. Today, we have mature algorithms for assessing morphological features at the cellular and tissue levels. In addition, there are promising efforts that link morphological features with biologically relevant information. While promising, these efforts mostly focus on narrow, well-defined questions. Developing a comprehensive report that is required in our setting requires an annotation strategy that captures all clinically relevant patterns specified in the WHO guidelines. Here, we propose and compare approaches aimed to balance the dataset and mitigate the biases in learning by automatically prioritising regions with clinical patterns underrepresented in the dataset. Our study demonstrates the opportunities active data enrichment can provide and results in a new lung-cancer dataset annotated to a degree that is not readily available in the public domain.