Artificial intelligence-based detection of stroke in non contrast computed tomography images

Background: Stroke is the second most leading cause of death, the World Health Organization defined stroke as 'rapidly developed clinical signs of focal (or global) disturbance of cerebral function, lasting more than 24 hours or leading to death, which is caused due to blockage or repture of brain blood vessel, out of which Intracranial hemorrhage (ICH) is most common.Recently, Artificial intelligence (AI) has shown great promise in the medical imaging domain, in which CT scan was a primary method for determining whether a stroke is ischemic or hemorrhagic.Also AI solution can serve as a prospective tool for non-contrast head CT scans to identify ICH and thus minimize false negatives Objectives: To explore and evaluate existing literature for detection of stroke in Non-contrast computerized tomography (NCCT) images and to implement and evaluate the deep learning model on NCCT images for detection of stroke. Material and Methods:The total of 2000 brain window CT images of stroke were collected online from the public dataset.The calculated sample size was 186 images with the precision level of ±7% where the confidence level was 95%.A Resnet-18 model was used as a deep learning model for the classification of intracranial hemorrhage and its types.Based on the manual screening of the images on the AI model, the highest sensitivity and specificity of the model were obtained at 70% for the training dataset and 30% for the testing dataset.On this Criterion of 70:30 the model was trained and tested using 131 (70%) and 55 (30%) images, respectively.Results: Out of 186 images 155 images were ICH-positive and 31 images were ICH-negative (normal) the deeplearning model achieved overall accuracy of 86% for Prediction of intracranial hemorrhage (ICH) and classifies its subtypes with Sensitivity of 86.2%, Specificity of 97.1, Error of 14.1% and Precision of 88%.Out of 186 images 178 images were were accurately labelled by the algorithm an eight were inaccurately labelled.SAH was the best and SDH was the least performed subtype out of the ICH-subtypes. Conclusion:The proposed method was able to accurately detect ICH and its subtypes like IVH, SDH, SAH, and EDH with high accuracy (86%), and was able to recognize the ICH-negative (Normal) suggesting it's potential for assisting radiologists and physicians in their clinical diagnosis workflow Our study demonstrates that, in a diverse clinical setting, an AI solution has the potential to assist radiologists and serve as a real-time clinical