SISDAL - Semantic Image Segmentation by Deep Active Learning in medical imaging applications

Without even noticing, artificial intelligence (AI) has brought innovations in many practical and commercial applications that become part of our everyday life. The success of AI was mainly due to disruptive shift of machine learning towards the use of cognitive deep learning systems, more specifically deep convolutional neural networks (CNNs). Bringing improvements to the hot topic of deep learning has become the main research interest of leading academic institutions, where also LBI-CFI and ICG are strongly contributing. As the world-leading IT companies, also the SME partner of our consortium tries to bring this cutting edge technology to customers that have little or no knowledge in artificial intelligence. One of the fundamental machine learning challenges in medical image analysis and computer vision is to automatically understand high-level semantic relationships between the objects of a depicted scene, which crucially depend on its component performing semantic image segmentation. Using CNNs for supervised semantic segmentation has brought unprecedented performance in medical image and computer vision applications that critically depend on success of accurate segmentation. However, training of deep neural networks crucially depends on a large amount of accurately annotated data. To establish such carefully annotated datasets, a huge amount of manual labelling work is required, an effort that is both time-consuming and costly. Therefore, for many companies, using the state-of-the-art machine learning technology is still not possible. Fortunately, for many real world problems, large amounts of unlabelled data are available for free. Promising dramatic reduction in the number of examples that a domain expert has to annotate without losing on performance of the machine learning method, active learning is putting a human expert in the loop between large amounts of unlabelled data and training deep networks. Active learning has not yet been used together with deep learning for semantic segmentation, since deep learning does not provide an uncertainty of predictions needed in the active learning loop to select new examples for annotation and large computational effort for re-training networks. Addressing these scientific questions together with technical challenges of combining active learning and deep learning, this project is the first to attempt bringing active learning closer towards commercial applications. The expected result of this project is a research prototype at TRL 4 for deep active learning in semantic image segmentation. On the long run this research prototype could be the foundation of a commercial product that reduces the time a human expert has to spend on annotation, thus significantly lowering the costs of bringing state-of-the-art machine learning methods to customers. Moreover, time to production could be significantly reduced, since only a small portion of large training datasets needs to be annotated. This may lead to companies using our developed deep active learning framework to enhance their competitiveness and attractiveness, potentially opening up new market possibilities.