Background: Accurate assessment of baseline infarct volume and the Alberta Stroke Program Early Computed Tomography Score (ASPECTS) on diffusion-weighted imaging (DWI) is crucial for treatment decision-making in candidates for mechanical thrombectomy (MT). However, evaluation of acute ischemic stroke lesions is affected by reader expertise, image quality, and interobserver variability. This study aimed to develop deep learning models for the automated assessment of infarct volume and ASPECTS on DWI in patients undergoing MT. Methods: We retrospectively analyzed patients with acute ischemic stroke due to large vessel occlusion who underwent emergent MT between September 2014 and December 2019. All input image masks were transferred to the SYNAPSE Creative Space, a cloud-based AI development platform (FUJIFILM Corporation). For semantic segmentation, acute DWI infarct volumes on 3D MRI were automatically segmented using a 3D U-Net model. For image classification, DWI-ASPECTS were automatically assessed using a convolutional neural network (CNN). The Dice score was calculated as (2 × overlapping area) / total area. Results: A total of 239 patients (152 male 64%, median age 75 years IQR 67–81, median NIHSS score 15 IQR 7–21, ICA: n=61, M1: n=95, M2: n=83) were included. The baseline median DWI infarct volume was 15.9 ml IQR 7.4–75.1, and the median DWI-ASPECTS was 6 IQR 4–8. Of the dataset, 187 DWI-MRIs were used for training and 52 for internal validation. Automatic infarct segmentation on 3D DWI-MRI demonstrated good accuracy (mean Dice score: 0.76). CNN-based automatic assessment of DWI-ASPECTS achieved a mean Dice score of 0.83. Conclusion: We developed automated models for infarct volume estimation and ASPECTS assessment on DWI in patients undergoing MT. Our findings suggest that a two-step approach—segmentation followed by classification—is feasible and may support physicians in treatment decision-making for MT candidates.
Shimoyama et al. (Thu,) studied this question.