Detection and classification of intracranial haemorrhage on CT images using a novel deep-learning algorithm.
Ji Young Lee, Jong Soo Kim, Tae Yoon Kim, Young Soo Kim
Author Information
Ji Young Lee: Department of Radiology, School of Medicine, Hanyang University Seoul Hospital, Seoul, Republic of Korea.
Jong Soo Kim: Institute for Software Convergence, Hanyang University, 222 Wangsimni-ro, Seongdong-gu, Seoul, 04763, Republic of Korea. jongkim57@hanyang.ac.kr.
Tae Yoon Kim: Department of Radiology, College of Medicine, Hanyang University Guri Hospital, Guri, Republic of Korea.
Young Soo Kim: Department of Neurosurgery, School of Medicine, Hanyang University Seoul Hospital, 222 Wangsimni-ro, Seongdong-gu, Seoul, 04763, Republic of Korea. ksy8498@hanyang.ac.kr.
A novel deep-learning algorithm for artificial neural networks (ANNs), completely different from the back-propagation method, was developed in a previous study. The purpose of this study was to assess the feasibility of using the algorithm for the detection of intracranial haemorrhage (ICH) and the classification of its subtypes, without employing the convolutional neural network (CNN). For the detection of ICH with the summation of all the computed tomography (CT) images for each case, the area under the ROC curve (AUC) was 0.859, and the sensitivity and the specificity were 78.0% and 80.0%, respectively. Regarding ICH localisation, CT images were divided into 10 subdivisions based on the intracranial height. With the subdivision of 41-50%, the best diagnostic performance for detecting ICH was obtained with AUC of 0.903, the sensitivity of 82.5%, and the specificity of 84.1%. For the classification of the ICH to subtypes, the accuracy rate for subarachnoid haemorrhage (SAH) was considerably excellent at 91.7%. This study revealed that our approach can greatly reduce the ICH diagnosis time in an actual emergency situation with a fairly good diagnostic performance.
