Artificial intelligence in scans for deciding quickly on stroke treatment
In stroke, fast diagnosis and treatment is key to prevent oxygen shortage from causing irreversible cell death. CT imaging has been the predominant method of choice for diagnosis. The AMC and Nico-lab have developed Artificial Intelligence (AI) methods for automated CT image analysis tasks to improve the stroke workup in clinical practice. As an alternative to CT, MR imaging gives an unprecedented view on the stroke brain, holding a strong promise for improving diagnosis and treatment selection. The development of AI methods for MR images will enhance the decision making process for stroke patients. We therefore aim to develop stroke treatment decision support by AI for analysis of MR images, in which Amsterdam UMC joins forces with Nico-lab.
Stroke is a major cardiovascular disease with severe impact on quality-of-life and huge costs for society of €45B for an expected 1.2 million patients in the EU in 2025. The introduction of intra-arterial treatment in 2015 has revolutionized stroke care, with positive outcome in 30-70% of patients.
AI-supported decision making in medical imaging traverses over a cascade of connected tasks which all need to be trained on large datasets. By simultaneously rather than independently learning these tasks, an increase in efficiency of 60-70% can be achieved, holding a great promise for stroke treatment decision support.
We developed a fast MRI scanning protocol with twelve-fold acceleration for stroke patients, and conducted a pilot study where we successfully acquired data in long-term follow-up stroke patients. We developed an AI algorithm for efficient reconstruction of highly accelerated MRI data. In international challenges, our algorithm has won in multiple categories of brain data and other organs. Furthermore, our algorithm was among the top performing methods when being used for data unseen during the training phase. We brought economy to deep learning by efficiently transferring knowledge between multiple successive tasks in medical image analysis, i.e., image reconstruction and segmentation of lesions in the brain. Furthermore, we were successful in segmenting lesions on CT after training on MRI, which was associated with outcome after acute ischemic stroke. We have developed a software platform for automated medical image data analysis over multiple tasks with consistency. This platform supports complex data analysis, in multiple computing environments, and is thus ready to be used in the cloud. An additional interdisciplinary code base was released with the journal of open source software. In benchmarking experiments, our algorithm was two times faster than the standard for MRI reconstruction. Of note, we had first optimized this benchmark method for its highest performance on Graphical Processing Units (GPUs). Still, our AI-network was two times faster, which we attribute to requiring fewer iterations to converge. We have published an online large dataset of raw MRI data in the cloud and made this dataset openly accessible. This will generate spin-off to other research projects.
