Introduction: Deep learning is a novel machine learning approach that enables automated extraction and classification of imaging features. We aimed to use deep learning to enhance detection of brain infarction.
Methods: Patients with acute ischemic stroke admitted to our institution between 2011-2015 were prospectively registered in an IRB-approved database. On a subset of these patients, radiologists annotated the images by marking infarct area. Data was split randomly into a training set and test set (80:20). Deep learning models, including a 3D multi-scale fully convolutional neural network, were developed and trained on the training set and independently tested on the test set. The performance of this model was compared to the expert consensus interpretation. Diagnostic test characteristics including the area under the curve (AUC) for the deep learning algorithm were calculated both at a voxel and imaging-study level. Computer-generated heat maps were created to denote the possibility of infarct.
Results: The study group included 114 patients (training set n = 92, test set n = 22). In the training set of 5,888 images, infarction was present in 602 (10.2%) images. In the testing set of 920 images, infarction was present in 130 (14.1%) images. A total of 1.5 billion voxels were used to train the model. The AUC for the deep learning algorithm for voxel accuracy was 0.973 (95% CI 0.972-0.974). Voxel accuracy, sensitivity, and specificity were 92%, 93%, and 92%, respectively. Positive predictive value (PPV) and negative predictive value (NPV) were 86% and 92%, respectively. The AUC for the deep learning algorithm for automated diagnosis of infarction at the imaging-study level was 0.91 (95% CI 0.90-0.94). Diagnostic accuracy, sensitivity, and specificity were 88%, 65%, and 91%, respectively. PPV and NPV were 49% and 95%, respectively.
Conclusions: A machine-learned algorithm employing a novel deep learning algorithm enabled accurate diagnosis of brain infarction.