AUTHOR=Wang Hao , Sun Yi , Zhu Jie , Zhuang Yuzhong , Song Bin 

TITLE=Diffusion-weighted imaging-based radiomics for predicting 1-year ischemic stroke recurrence

JOURNAL=Frontiers in Neurology

VOLUME=Volume 13 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/neurology/articles/10.3389/fneur.2022.1012896

DOI=10.3389/fneur.2022.1012896

ISSN=1664-2295

ABSTRACT=Purpose: To investigate radiomics based on DWI (diffusion-weighted imaging) for predicting one-year ischemic stroke recurrence. 
Methods: A total of 1580 ischemic stroke patients were enrolled in this retrospective study conducted from January 2018 to April 2021. Demographic and clinical characteristics were compared between recurrence and non-recurrence groups. On DWI, lesions were segmented using a 2D U-Net automatic segmentation network. Further radiomics features extraction was done using the segmented mask matrix on DWI and the corresponding ADC map. Moreover, radiomics features were extracted. A sparse representation feature selection model was performed to select features. A recurrent neural network (RNN) was used to further classification. The area under the receiver operating characteristic curve values for model performance were obtained.
Results: A total of 1003 ischemic stroke patients were included in the final analysis. 85 patients (8.5%) recurred in one year, patients in recurrence group were older than non-recurrence group (P = 0.003). Stroke subtype was significantly different between recurrence and non-recurrence groups, and cardioembolic stroke (11.3%) and large artery atherosclerosis patients (10.3%) showed higher recurrence percentage (P = 0.005). Medications after discharge were found significantly different between the two groups (P = 0.004). The area under the curve (AUC) of clinical-based model and radiomics-based model were 0.675 and 0.779, respectively. With an AUC of 0.847, the model that combined clinical and radiomic characteristics performed better.
Conclusion: DWI-based radiomics could help to predict one-year ischemic stroke recurrence. Sparse representation feature selection and RNN classification could contribute to our robust prediction performance.