Segmenting Hemorrhagic and Ischemic Infarct Simultaneously From Follow-Up Non-Contrast CT Images in Patients With Acute Ischemic Stroke

Abstract

Cerebral infarct volume (CIV) measured from follow-up non contrast CT (NCCT) scans of acute ischemic stroke (AIS) patients is an important radiologic outcome measure of the effectiveness of ischemic stroke treatment. Post-treatment CIV in NCCT of AIS patients typically includes ischemic infarct only. In around 10% of AIS patients, however, hemorrhagic transformation or frank hemorrhage occurs along with ischemic infarction. Manual segmentation used to segment CIV into these two components in clinical practice is tedious and user dependent. Although automated segmentation methods exist, they can only segment either hemorrhage or ischemic infarct alone. In order to measure post-treatment CIV more efficiently, a novel joint segmentation approach is proposed to segment ischemic and hemorrhage infarct simultaneously. The proposed method makes use of advances in deep learning and convex optimization techniques. Specifically, convolutional neural network learned semantic information, local image context, and high-level user initialized prior are integrated into a multi-region time-implicit contour evolution scheme, which can be globally optimized by convex relaxation. The proposed segmentation approach is quantitatively evaluated using 30 patient images using Dice similarity coefficient and the mean and maximum absolute surface distance, compared to the gold standard of manual segmentation. The results show that the proposed semi-automatic segmentation is accurate and robust, outperforming some state-of-the-art semi-and automatic segmentation approaches.