Precise diagnosis of intracranial hemorrhage and subtypes using a three-dimensional joint convolutional and recurrent neural network

• DOI: 10.1007/s00330-019-06163-2
• Publication Date: 2019-04-30T18:33:50Z
• Authors: Hai Ye; Feng Gao; Youbing Yin; Danfeng Guo; Pengfei Zhao; Yi Lu; Xin Wang; Junjie Bai; Kunlin Cao; Qi Song; Heye Zhang; Wei Chen; Xuejun Guo; Jun Xia
• Keywords: Neuroradiology; Interventional radiology

To evaluate the performance of a novel three-dimensional (3D) joint convolutional and recurrent neural network (CNN-RNN) for detection intracranial hemorrhage (ICH) its five subtypes (cerebral parenchymal, intraventricular, subdural, epidural, subarachnoid) in non-contrast head CT.A total 2836 subjects (ICH/normal, 1836/1000) from three institutions were included this ethically approved retrospective study, with 76,621 slices CT scans. ICH annotated by independent experienced radiologists, majority voting as reference standard both subject level slice level. Ninety percent data was used training validation, rest 10% final evaluation. A CNN-RNN classification framework proposed, flexibility to train when subject-level or slice-level labels are available. The predictions compared interpretations junior radiology trainees an additional senior radiologist.It took our algorithm less than 30 s on average process 3D scan. For two-type task (predicting bleeding not), achieved excellent values (≥ 0.98) across all reporting metrics five-type subtypes), > 0.8 AUC subtypes. generally superior tasks.The proposed method able accurately detect fast speed, suggesting potential assisting radiologists physicians their clinical diagnosis workflow.• deep learning developed subtype classification, which has either labels. • This is accurate at detecting automated work, reduce initial misinterpretations.