Abstract
BACKGROUND: CT is commonly used to image patients with ischaemic stroke but radiologist interpretation may be delayed. Machine learning techniques can provide rapid automated CT assessment but are usually developed from annotated images which necessarily limits the size and representation of development data sets. We aimed to develop a deep learning (DL) method using CT brain scans that were labelled but not annotated for the presence of ischaemic lesions. METHODS: We designed a convolutional neural network-based DL algorithm to detect ischaemic lesions on CT. Our algorithm was trained using routinely acquired CT brain scans collected for a large multicentre international trial. These scans had previously been labelled by experts for acute and chronic appearances. We explored the impact of ischaemic lesion features, background brain appearances and timing of CT (baseline or 24-48 hour follow-up) on DL performance. RESULTS: From 5772 CT scans of 2347 patients (median age 82), 54% had visible ischaemic lesions according to experts. Our DL method achieved 72% accuracy in detecting ischaemic lesions. Detection was better for larger (80% accuracy) or multiple (87% accuracy for two, 100% for three or more) lesions and with follow-up scans (76% accuracy vs 67% at baseline). Chronic brain conditions reduced accuracy, particularly non-stroke lesions and old stroke lesions (32% and 31% error rates, respectively). CONCLUSION: DL methods can be designed for ischaemic lesion detection on CT using the vast quantities of routinely collected brain scans without the need for lesion annotation. Ultimately, this should lead to more robust and widely applicable methods.