Deep Learning disconnectomes to accelerate and improve long-term predictions for post-stroke symptoms

Abstract

Deep learning as a truly transformative force is revolutionizing a wide range of fields, making a significant difference in medical imaging, where recent advancements have yielded some truly remarkable outcomes. In a connected brain, maps of white matter damage -- otherwise known as disconnectomes -- are essential for capturing the effects of focal lesions. However, the current tools for obtaining such information are prohibitively slow and not admitted for clinical usage. Here, we have explored the potential of deep-learning models to accurately generate disconnectomes in a population of stroke survivors. We trained a 3D U-Net algorithm to produce deep-disconnectomes from binary lesion masks. This artificial neural network was able to capture most information obtained in conventional disconnectomes, i.e., statistical maps filtering normative white-matter networks, but output a deep-disconnectome 170 times faster - compared to disconnectome computation with the state-of-the-art BCBToolkit software. Moreover, the deep-disconnectomes were challenged to predict cognitive and behavioral outcomes one-year post-stroke. In an additional cohort of N=139 stroke survivors, N=86 neuropsychological scores were predicted from deep-disconnectomes achieving, on average, 85.2% of accuracy and R2= 0.208. The deep-disconnectomes predictivity power outperformed the conventional disconnectome predictions for clinical scores. In summary, we have achieved a significant milestone for clinical neuroimaging by accelerating and ameliorating the creation of disconnectome maps using deep learning. By integrating deep learning into the management of stroke, one of the most prevailing catalysts for acquired disabilities, we deepen our understanding of its impact on the brain. This novel approach may offer potential avenues for acute intervention, ultimately enhancing patients overall quality of life.