Abstract
BackgroundSpontaneous intracranial hemorrhages are life-threatening conditions that require fast and accurate diagnosis. We hypothesized that deep learning (DL) could be utilized to detect these hemorrhages with a high accuracy.MethodsWe developed a DL solution for detecting spontaneous intracerebral (ICH), intraventricular (IVH) and subarachnoid hemorrhages (SAH) from head non-contrast CT (NCCT) scans. The solution included four convolutional neural network (CNN) base models for different hemorrhage types and a CNN metamodel that was trained on top of the base models. We validated the performance of the solution by using a retrospective real-world dataset of consecutive emergency head NCCTs imaged during a 3-month period in 10 different hospitals. The head NCCTs with hemorrhages were stratified into groups by delay from symptom onset to NCCT imaging to better evaluate the suitability of the solution for emergency use.ResultsThe real-world validation dataset included 7797 emergency head NCCTs that were imaged between October 1stand December 31st2021. Of these, 118 were reported to show spontaneous intracranial hemorrhages by on-call radiologists, and 7679 were reported negative for hemorrhages. The developed solution detected all reported 78 (sensitivity 100%) spontaneous intracranial hemorrhages if the head NCCT was presumably or confirmedly taken within 12 hours of symptom onset. When assessed for hemorrhages imaged 12 to 24 hours after symptom onset (13 cases), the sensitivity was 76.5 %. Overall sensitivity for detecting spontaneous intracranial hemorrhages on head NCCTs that were imaged with any delay from symptom onset was 89.8 %, and specificity was 89.5 %. The solution also detected five cases that were missed by on-call radiologists.ConclusionsThe DL solution showed high sensitivity for detecting spontaneous ICHs, IVHs and SAHs within the same time window in which also modern CT scanners work best for detecting acute blood on head NCCTs.
Publisher
Cold Spring Harbor Laboratory