Image analysis in pathology is one of the most challenging visual tasks performed by humans. Finding tumor cells in large tissue slides is often like searching for a needle in a haystack. This challenge gets even harder when building systems that scale to the real world: what a pathology probe looks like can be different from lab to lab due to differences in preprocessing and stainings.
We solve this challenge by building systems based on artificial intelligence (AI) that work out of the box. Our BreastIHC is the first ever commercially available product that is able to detect tumor cells in arbitrary IHC-stained breast images without the need for manual fine-tuning.
We have developed proprietary clinical-grade deep learning architectures that are tailored to pathology and achieve results on par with experts. Combining human experts with our AI leads to accuracies beyond current standards.
Our team has already built secure, fast and robust AI systems for millions of users. We know what it takes to scale AI from the benign R&D setting to the noisy real world and systems with high regulatory demands in clinical diagnostics.
Image analysis tools need to earn the trust of the experts. Modern methods in explainable AI help to open the black box and are an essential component to achieve this trust. We use such methods to provide the needed guidance to pathologists and clinicians.
In addition to labeled data, there is a wealth of images without annotations in pathology. We harvest this treasure and learn relevant patterns even from unannotated data by using specialized methods such as self-supervised learning and generative adversarial networks.
We are lead author of the German “Guideline for the development of deep learning based image analysis systems” by the German Institute for Standardization (DIN) (“Leitfaden für die Entwicklung von Deep-Learning-Bilderkennungssystemen”). Currently, we are preparing the follow-up guideline that focuses on image analysis in medicine.