While the number of stroke patients is increasing worldwide and every fifth stroke survivor is developing long-term cognitive impairment, its prediction becomes more and more important. In this work, we address the challenge of predicting any long-term cognitive impairment after a stroke using deep learning. We explore multi-task learning that combines the cognitive classification with the segmentation of brain lesions such as infarct and white matter hyperintensities or the reconstruction of the brain. Our approach is further expanded to include clinical non-imaging data to the input imaging information. The multitask model using an autoencoder for reconstruction achieved the highest performance in classifying post-stroke cognitive impairment when only imaging data is used. The performance can be further improved by incorporating clinical information using a previously proposed dynamic affine feature map transformation. We developed and tested our approach on an in-house acquired dataset of magnetic resonance images specifically used to visualize stroke damage right after stroke occurrence. The patients were followed-up after one year to assess their cognitive status. The multi-task model trained on infarct segmentation on diffusion tensor images and enriched with clinical non-imaging information achieved the best overall performance with a balanced accuracy score of 70.3% and an area-under-the-curve of 0.791.     «

While the number of stroke patients is increasing worldwide and every fifth stroke survivor is developing long-term cognitive impairment, its prediction becomes more and more important. In this work, we address the challenge of predicting any long-term cognitive impairment after a stroke using deep learning. We explore multi-task learning that combines the cognitive classification with the segmentation of brain lesions such as infarct and white matter hyperintensities or the reconstruction of th...     »