ACUTE ISCHAEMIC STROKE (AIS) LESION DETECTION WITH A CONVOLUTIONAL DEEP LEARNING MODEL (ID 708)
Abstract
Background And Aims
In this pilot study, we used deep learning to detect acute ischaemic stroke (AIS) lesions on brain CT. Our aims were to classify the side of brain affected by AIS and to identify features used for classification.
Methods
We built a 7-layer convolutional neural network (CNN) convolving on 11 uniformly sampled axial slices per scan before averaging each feature across all slices. We sampled left and right hemispheres separately and thus controlled for most potential confounders as every patient provided both positive and negative cases. We processed 2841 axial CTs from 2350 patients with expert labelling (presence/absence of AIS lesions) from The Third International Stroke Trial. Images were acquired at baseline, 24-48 hours follow-up and when additional clinical need arose. 2001 scans from 1063 patients were utilised for training, 429 scans from another 229 patients for validation and 411 scans from different 229 patients for testing.
Results
The CNN achieved 76% classification accuracy overall, 79% on follow-up scans versus 70% on baseline scans i.e., lesions are more identifiable at later times (p=0.0033). Training on single middle brain (~basal ganglia) slices achieved 70% accuracy versus 65% on upper middle brain (~lateral ventricles) and 55% on lower middle brain (~posterior fossa), consistent with common AIS lesion locations. Accuracy remained at 70% when individual inputs were resized to 1/8 original height and width (i.e. when lesions should be less visible).
Conclusions
Our deep learning model had reasonable accuracy for detecting AIS lesion side on CT but may use global brain structure and lesion locations for classification.
Trial Registration Number
Not applicable
