# Automated quantification of the Alberta Stroke Programme Early CT Score (ASPECTS) and infarct core on diffusion-weighted imaging in acute ischaemic stroke: multi-centre validation and severity stratification

## TL;DR

A deep learning model was developed to automatically assess stroke severity using brain scans, showing strong agreement with expert evaluations and potential for clinical use.

## Contribution

A novel deep learning framework for automated ASPECTS scoring and infarct core volume measurement in acute stroke, validated across multiple centers.

## Key findings

- The 3D U-Net model achieved high correlation with manual segmentation (Dice coefficient of 0.801 and Spearman correlation of 0.988).
- Automated ASPECTS scores showed strong agreement with expert ratings in both internal and external test sets.
- Severity stratification thresholds for mild and moderate–severe stroke groups were identified as 6 for ASPECTS and 27.86 mL for core volume.

## Abstract

The purpose of this study was to develop and validate a deep learning framework for simultaneous automated quantification of diffusion-weighted imaging-the Alberta Stroke Programme early CT score (DWI-ASPECTS) and infarct core volume in middle cerebral artery acute ischaemic stroke (MCA-AIS) and to evaluate its clinical utility for severity stratification in multi-centre settings. A cohort of 738 patients diagnosed with MCA-AIS from four centres was divided into a train set (n = 408), a validation set (n = 116), an internal test set (n = 60) and an external test set (n = 154). A 3D U-Net architecture was trained for simultaneous infarct segmentation and the ASPECTS region analysis. Model performance was compared against expert neuroradiologists using intraclass correlation coefficients and the Spearman correlation coefficient. Furthermore, we investigated the correlation between ASPECTS deduction frequency, ASPECTS score, core volume and the severity of MCA-AIS. The 3D U-Net model showed a high correlation with manual segmentation, achieving a Dice coefficient of 0.801 and a Spearman correlation coefficient of 0.988 for volume measurements in the external test set. In both the internal and external test sets, the automated ASPECTS by the deep learning (DL) model (Automated), manual ASPECTS by raters (Raters) and manual ASPECTS by raters on DWI images registered with the template (Raters_template) exhibited a strong correlation and excellent agreement. The cortical regions (M1–M6) were particularly relevant in patients classified into the moderate–severe group. The threshold values for the mild group and moderate–severe group on receiver operating characteristic curve analysis for DWI-ASPECTS and Core volume, were 6 and 27.86 mL, respectively. The DL model demonstrated comparable performance to neuroradiologists’ evaluation, potentially serving as an ancillary tool for physicians in making urgent clinical decisions. The severity of MCA-AIS was significantly associated with the specific ASPETCS regions and core volume, which may aid in identifying moderate–severe MCA-AIS.

Wei et al. report that a novel deep learning model fully automates the Alberta Stroke Program Early CT Score (ASPECTS) scoring and infarct core volume measurement from diffuse-weighted imaging (DWI) scans. This tool was validated on multi-centre data and may aid in rapid stroke assessment.

Graphical AbstractFor image description, please refer to the figure legend and surrounding text.