Abstract
Background: Ischemic stroke is the most common type of stroke. In the context of ischemic stroke imaging, automated segmentation methods facilitate deeper exploration of imaging data by providing consistent measurements and quantitative analyses, thereby reducing human errors while enhancing physician efficiency. This study presents a specific type of convolutional neural network (NN) to automate the segmentation of ischemic stroke lesions.
Methods: A deep NN called U-Net was developed for the automatic segmentation of ischemic stroke lesions using a U-shaped architecture. Then, the He-normal initialization algorithm was employed to address the issues of gradient explosion or vanishing gradients.
Results: The U-Net model was evaluated on a separate dataset consisting of samples collected from Tabriz University of Medical Sciences, which included 837 images of patients with acute ischemic stroke who sought treatment interventions at the center during 2021 and underwent magnetic resonance imaging (MRI). The obtained Dice coefficient (DC) values varied, depending on different images and their complexities (0.89–0.98). Notably, most images exhibited an average DC of 0.96, with approximately 70% of the images achieving a value of 0.96. Moreover, the annotation time for the network was reported to be 7 seconds for 10 images, while a radiologist took approximately 4 minutes for the same set of images.
Conclusion: Overall, the U-Net algorithm could enhance the speed and quality of service delivery, thereby reducing the time for patient care. Consequently, this improvement is likely to result in better patient outcomes and a decrease in stroke-associated complications.