# Development of a Deep Learning Method to Identify Acute Ischemic Stroke Lesions on Brain CT

**Authors:** Alessandro Fontanella, Wenwen Li, Grant Mair, Antreas Antoniou, Eleanor Platt, Paul Armitage, Emanuele Trucco, Joanna Wardlaw, Amos Storkey  
**arXiv:** [2309.17320](https://arxiv.org/abs/2309.17320)  
**Date:** 2023-09-29  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2309.17320v1/

## Key points

- Computed Tomography (CT) is commonly used for imaging acute ischemic stroke (AIS) patients
- Interpretation of CT scans by radiologists can be time-consuming and subject to inter-observer variability
- Deep learning (DL) techniques can provide automated CT brain scan assessment
- DL methods usually require annotated images, but this study used labelled but not annotated CT brain scans from AIS patients
- A convolutional neural network-based DL algorithm was developed using routinely-collected CT brain scans from the Third International Stroke Trial (IST-3)
- The DL model was designed to detect AIS lesions and classify the side of the brain affected
- The accuracy of the DL method for detecting lesion presence and classifying the side of the brain affected was 72%
- Accuracy improved for larger lesions or multiple lesions, reaching 80% for larger lesions and 87% for two lesions or more
- Follow-up scans had higher accuracy compared to baseline scans, with 76% accuracy for follow-up scans and 67% accuracy for baseline scans
- Chronic brain conditions reduced accuracy, particularly non-stroke lesions and old stroke lesions with error rates of 32% and 31%, respectively
- This study demonstrates that DL methods can be designed for AIS lesion detection on CT using routinely-collected data without strict research protocols
- DL methods have potential for more robust and widely applicable clinical applications in stroke imaging.

## Summary

Computed Tomography (CT) is commonly used to image acute ischemic stroke (AIS) patients and its interpretation by radiologists can be time-consuming and subject to inter-observer variability. Deep learning (DL) techniques can provide automated CT brain scan assessment but usually require annotated images. In this study, the researchers developed a DL method for AIS using labelled but not annotated CT brain scans from patients with AIS. They designed a convolutional neural network-based DL algorithm using routinely-collected CT brain scans from the Third International Stroke Trial (IST-3), which were not acquired using strict research protocols. The DL model was developed to detect AIS lesions and classify the side of the brain affected. The researchers explored how AIS lesion features, background brain appearances, and timing impacted DL performance. They analyzed 5772 unique CT scans of 2347 AIS patients with a median age of 82; according to expert labeling, 54% of the scans had visible AIS lesions. The best-performing DL method achieved 72% accuracy for detecting lesion presence and classifying the side of the brain affected. The accuracy improved for larger lesions (80%) or multiple lesions (87% for two lesions and 100% for three or more). Follow-up scans had higher accuracy (76%) compared to baseline scans (67%). However, chronic brain conditions reduced accuracy particularly non-stroke lesions and old stroke lesions with error rates of 32% and 31%, respectively. This study demonstrates that DL methods can be designed for AIS lesion detection on CT using routinely-collected data without strict research protocols. The findings suggest that these methods have potential for more robust and widely applicable clinical applications in stroke imaging.