Robust and Adaptive AI for Digital Pathology

The digitization of pathology opens up a wide field of applications that can be supported by AI-based analysis like the detection of tumors or a quantitative assessment of tissue composition. This contribution demonstrates possible ways on how to approach challenges in digital pathology like the robustness against data heterogeneity or the detection of out-of-distribution data. Moreover, the principle of prototypical few-shot models is explained, which can be adapted to new tasks with only a few labeled examples without any retraining of the underlying model parameters. In this chapter we show the suitability of a prototypical few-shot classification model for tumor detection in two different organs and a prototypical few-shot segmentation model for tumor composition analysis. Finally, a workflow for the creation of a dedicated AI model by only providing a few annotations within the MIKAIA^®software of Fraunhofer IIS is presented.