Treatment selection is becoming increasingly more important in acute ischemic stroke patient care. Clinical variables and radiological image biomarkers (old age, pre-stroke mRS, NIHSS, occlusion location, ASPECTS, among others) have an important role in treatment selection and prognosis. Radiological biomarkers require expert annotation and are subject to inter-observer variability. Recently, Deep Learning has been introduced to reproduce these radiological image biomarkers. Instead of reproducing these biomarkers, in this work, we investigated Deep Learning techniques for building models to directly predict good reperfusion after endovascular treatment (EVT) and good functional outcome using CT angiography images. These models do not require image annotation and are fast to compute. We compare the Deep Learning models to Machine Learning models using traditional radiological image biomarkers. We explored Residual Neural Network (ResNet) architectures, adapted them with Structured Receptive Fields (RFNN) and auto-encoders (AE) for network weight initialization. We further included model visualization techniques to provide insight into the network's decision-making process. We applied the methods on the MR CLEAN Registry dataset with 1301 patients. The Deep Learning models outperformed the models using traditional radiological image biomarkers in three out of four cross-validation folds for functional outcome (average AUC of 0.71) and for all folds for reperfusion (average AUC of 0.65). Model visualization showed that the arteries were relevant features for functional outcome prediction. The best results were obtained for the ResNet models with RFNN. Auto-encoder initialization often improved the results. We concluded that, in our dataset, automated image analysis with Deep Learning methods outperforms radiological image biomarkers for stroke outcome prediction and has the potential to improve treatment selection.
- Acute ischemic stroke
- Radiological images
- Deep learning
- Structured receptive fields
- Gradient-weighted class activation mapping