The use of artificial intelligence in diagnostic medical imaging has shown impressive accuracy and sensitivity in the identification of imaging abnormalities. Assisting healthcare professionals with detection, segmentation and classification, it can be of immense value for screenings and precision medicine. Our AI for Health engineers used Generative Adversarial Networks to segment X-rays of lungs to detect Covid-19. They applied U-nets and mask-RCNN for instance segmentation of cervical vertebrae to predict patient developing hernia.

Medical AI is teeming with sensitive patient data used to train machine learning models. We teach our AI engineers how to handle sensitive information without compromising the model accuracy. We deploy techniques like Federated Learning to train algorithms across multiple decentralized servers holding local data without sharing them.

For transparent algorithmic decision making we deploy the techniques of Explainable AI or sometimes called Trustworthy AI. In healthcare, it’s important to see how the machine learning model arrived at a certain decision. How does it assess which preterm baby is at risk of getting sepsis or what does the neural network pay attention to when segmenting a medical image (as seen on our heatmap of a lung scan)?

We collaborate with medical professionals to deploy state-of-the-art medical analysis in our Challenges. In order to develop a usable AI solution, the participants need to understand the relevant hardware and the medical procedures. To improve hernia detection models for example, it’s of great value to have a neuroscientist explain the principles of radiology imaging, hernia diagnosis and its treatment to Challenge participants. Or the analysis of electrocardiograms to prevent heart failure.