Prediction of response to thrombolysis in acute stroke using neural network analysis of CT perfusion imaging

Abstract

Background: In ischaemic stroke patients undergoing reperfusion therapy, the amount of salvageable tissue, that is, extent of the ischaemic penumbra, predicts the clinical outcomes. CT perfusion (CTP) enables quantification of penumbral tissues to guide decision making, and current programmes have automated its analysis. More advanced machine learning techniques utilising the CTP maps may improve prediction beyond the ischaemic volume measures.

Method: We determined whether applying convolutional neural networks (CNN), a key machine learning technique in modelling image-label relationships, to post-processed CTP maps improved prediction of outcome, assessed by 3 months modified Rankin scale (mRS). Patients who underwent thrombolysis but not thrombectomy were included. CTP maps of a retrospective cohort of 230 patients with middle cerebral artery stroke were used to develop the model, which was validated in an independent cohort of 129 patients.

Results: We constructed a CNN model that predicted a favourable post-thrombolysis outcome (mRS 0-2 at 3 months) with an area under receiver-operator characteristics curve (AUC) of 0.792 (95% CI, 0.707-0.877). This model outperformed a currently clinically used MISTAR software using previously validated thresholds (AUC = 0.583, 95% CI, 0.480-0.686) and a model modified using thresholds from the derivation cohort (AUC = 0.670, 95% CI, 0.571-0.769). By combining CNN-derived features and baseline demographic features, the prediction AUC was improved to 0.865 (95% CI, 0.794-0.936).

Conclusion: CNN improved prediction of post-thrombolysis outcome, and may be useful in selecting which patients benefit from thrombolysis.

Keywords: CT perfusion; Machine learning; neural network; thrombolysis.

Figure 1.

Patient selection flow chart. The derivation cohort consists of both the training and validation datasets.

Figure 2.

Correlation between clinical and CTP features and 3 months mRS in descending order. Correlation was performed using the Spearman test. The p values are adjusted for multiple testing using Bonferroni correction. p values were log transformed such that log p value of 0, 1, 2 and 3 corresponds to p values of 1, 0.1, 0.01 and 0.001 respectively. Abbreviations: IHD: ischaemic heart disease; AF: atrial fibrillation; HTN: hypertension; HC: hypercholesterolaemia; HF: heart failure; Core: ischaemic core volume; Smoker: current and/or previous smoking history.

Figure 3.

Saliency mapping of the CTP regions that most strongly activate the CNN. The cerebral blood flow (CBF), cerebral blood volume (CBV) and the saliency map (saliency) from three patients in the replication cohort were shown (P1, P2 and P3). The values in each map were rescaled to between 0 and 1 (minimum and maximum values in a volume of a CTP map respectively). In the saliency maps, regions with higher values indicate their higher degree of importance in contributing towards the CNN output.