ABSTRACT
Background: We compared the ischemic core and hypoperfused tissue volumes estimated by RAPID and JLK-CTP, a newly developed automated computed tomography perfusion (CTP) analysis package. We also assessed agreement between ischemic core volumes by two software packages against early follow-up infarct volumes on diffusion-weighted images (DWI). Methods: This retrospective study analyzed 327 patients admitted to a single stroke center in Korea from January 2021 to May 2023, who underwent CTP scans within 24 h of onset. The concordance correlation coefficient (ρ) and Bland-Altman plots were utilized to compare the volumes of ischemic core and hypoperfused tissue volumes between the software packages. Agreement with early (within 3 h from CTP) follow-up infarct volumes on diffusion-weighted imaging (n = 217) was also evaluated. Results: The mean age was 70.7 ± 13.0 and 137 (41.9%) were female. Ischemic core volumes by JLK-CTP and RAPID at the threshold of relative cerebral blood flow (rCBF) < 30% showed excellent agreement (ρ = 0.958 [95% CI, 0.949 to 0.966]). Excellent agreement was also observed for time to a maximum of the residue function (T max) > 6 s between JLK-CTP and RAPID (ρ = 0.835 [95% CI, 0.806 to 0.863]). Although early follow-up infarct volume showed substantial agreement in both packages (JLK-CTP, ρ = 0.751 and RAPID, ρ = 0.632), ischemic core volumes at the threshold of rCBF <30% tended to overestimate ischemic core volumes. Conclusion: JLK-CTP and RAPID demonstrated remarkable concordance in estimating the volumes of the ischemic core and hypoperfused area based on CTP within 24 h from onset.
ABSTRACT
Background and purpose: Multiple attempts at intracranial hemorrhage (ICH) detection using deep-learning techniques have been plagued by clinical failures. We aimed to compare the performance of a deep-learning algorithm for ICH detection trained on strongly and weakly annotated datasets, and to assess whether a weighted ensemble model that integrates separate models trained using datasets with different ICH improves performance. Methods: We used brain CT scans from the Radiological Society of North America (27,861 CT scans, 3,528 ICHs) and AI-Hub (53,045 CT scans, 7,013 ICHs) for training. DenseNet121, InceptionResNetV2, MobileNetV2, and VGG19 were trained on strongly and weakly annotated datasets and compared using independent external test datasets. We then developed a weighted ensemble model combining separate models trained on all ICH, subdural hemorrhage (SDH), subarachnoid hemorrhage (SAH), and small-lesion ICH cases. The final weighted ensemble model was compared to four well-known deep-learning models. After external testing, six neurologists reviewed 91 ICH cases difficult for AI and humans. Results: InceptionResNetV2, MobileNetV2, and VGG19 models outperformed when trained on strongly annotated datasets. A weighted ensemble model combining models trained on SDH, SAH, and small-lesion ICH had a higher AUC, compared with a model trained on all ICH cases only. This model outperformed four deep-learning models (AUC [95% C.I.]: Ensemble model, 0.953[0.938-0.965]; InceptionResNetV2, 0.852[0.828-0.873]; DenseNet121, 0.875[0.852-0.895]; VGG19, 0.796[0.770-0.821]; MobileNetV2, 0.650[0.620-0.680]; p < 0.0001). In addition, the case review showed that a better understanding and management of difficult cases may facilitate clinical use of ICH detection algorithms. Conclusion: We propose a weighted ensemble model for ICH detection, trained on large-scale, strongly annotated CT scans, as no model can capture all aspects of complex tasks.
