Deep learning-based identification and localization of intracranial hemorrhage in patients using a large annotated head computed tomography dataset: A retrospective multicenter study

Abstract

Background

Accurately identifying and localizing the five subtypes of intracranial hemorrhage (ICH) are crucial steps for subsequent clinical treatment; however, the lack of a large computed tomography (CT) dataset with annotations of the categorization and localization of ICH considerably limits the development of deep learning-based identification and localization methods. We aimed to construct this large dataset and develop a deep learning-based model to identify and localize the five ICH subtypes, including intraventricular hemorrhage (IVH), intraparenchymal hemorrhage (IPH), subdural hemorrhage (SDH), subarachnoid hemorrhage (SAH), and epidural hemorrhage (EDH), in non-contrast head CT scans.

Methods

Based on the public Radiological Society of North America (RSNA) 2019 dataset, we constructed a large CT dataset named RSNA 2019+ that was annotated for bleeding localization of the five ICH subtypes by three radiologists. An improved YOLOv8 architecture with the bidirectional feature pyramid network was proposed and trained using the RSNA 2019+ training dataset and evaluated on the RSNA 2019+ test dataset. The public CQ500, and two private datasets collected from the Xinqiao and Sunshine Union Hospitals, respectively, were also annotated to perform multicenter validation. Furthermore, the performance of the deep learning model was compared with that of four radiologists. Multiple performance metrics, including the average precision (AP), precision, recall and F1-score, were used for performance evaluation. The McNemar and chi-squared tests were performed, and the 95% Wilson confidence intervals were given for the precision and recall.

Results

There were 175,125; 4,707; 8,259; and 3,104 bounding boxes after annotation on the RSNA 2019+; CQ500+; and the PD 1 and PD 2 datasets, respectively. With an intersection-over-union threshold of 0.5, the APs of IVH, IPH, SAH, SDH and EDH are 0.852, 0.820, 0.574, 0.639, and 0.558, respectively, yielding a mean average precision (mAP) of 0.688 for our proposed deep learning model on the RSNA 2019+ test dataset. For the multicenter validation involving the three external datasets, the mAPs for CQ500, PD1, and PD2 were 0.594, 0.734, and 0.66, respectively, which is comparable to those of radiologist with eight years of experience in head CT interpretation.

Conclusion

The deep learning model developed from the constructed RSNA 2019+ dataset exhibited good potential in identifying and localizing the five ICH subtypes in CT slices and has the potential to assist in the clinical diagnosis.