Recently deep learning artificial intelligence technologies have started to emerge in the medical computer vision domain and, encouraged by the unprecedented amount of data in the health domain, lead to state-of-the-art models for, among others, identifying skin cancer and detecting lesions. However, despite their large potential of reducing medical costs and improving quality of care, the widespread adaption of these models is still lacking due to severe limitations. The key problems are their inability to utilise unlabelled data, the theoretical understanding of their training mechanisms, and a lack of interpretability of their decision-making processes. MedEx will develop novel interpretable techniques that can learn in the presence of little labeled data and that can be applied across hospitals. Further, the project will develop theoretical understanding of the training mechanisms in these models and produce frameworks to analyze them. The techniques will be evaluated on a range of medical image modalities and tasks, with specific focus on tumor detection and detection of diabetic retinopathy. MedEx is an ambitious project, which is enabled through an interdisciplinary consortium consisting of both national and international experts in artificial intelligence and the medical field. The anticipated scientific solutions will have a reach far beyond the health sector and contribute to moving the state-of-the-art within the field of artificial intelligence.
A large fraction of the data gathered in health care consists of images and as the field transitions towards data-driven health, the opportunity for clinical decision support systems and thus the need for processing medical images becomes apparent. Deep learning artificial intelligence technologies have emerged as the state-of-the-art for image processing as they achieve unprecedented predictive performance due to their ability to learn complex representations from the input data. However, in safety-critical domains, such as the medical domain, there are key-obstacles that need to be resolved before wide spread adaption in the clinic. Obstacles for Universal Healthcare are the inability of deep learning approaches to provide interpretable solutions, the lack of theoretical understanding of deep learning models, and the lack of annotated patients in the health domain. MedEx will develop interpretable data-efficient methods that are able to use unannotated data and are applicable across hospitals and not limited to a particular imaging protocol. The primary application focus is the detection of lung-cancer from PET/MR/CT images in order to improve early-detection and thereby improve patient survivability and quality of care. The secondary application is diabetic retinopathy detection from fungus images to enable efficient systematic screening. While methodological advances are generic and will impact data-driven health and science beyond, concrete results and outcomes will find application through MedEx's clinical stakeholders. Despite being a high-risk endeavor, the project is feasible due to the combination of a high-quality team, top international collaborators, and direct involvement of clinical stakeholders.