Perfusion imaging is very important for early assessment of strokes due to its ability to measure blood flow, transition times and dispersion. Deep learning approaches are able to perform automatic segmentations, but have had limited accuracy so far for clinically important small lesions. We present an extension to the popular U-Net architecture that concatenates higher-level features and improves their propagation throughout the network using additional skip connections. The new method is evaluated on public perfusion datasets and achieves substantially improved accuracy (33% lower surface distance than U-Net) and robustness in particular for smaller stroke lesions that are difficult to detect.
|Title of host publication||2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018)|
|Number of pages||4|
|ISBN (Print)||978-1-5386-3635-0, 978-1-5386-3637-4|
|Publication status||Published - 23.05.2018|
|Event||2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018) - Washington, DC, United States|
Duration: 04.04.2018 → 07.04.2018