Strengthening deep-learning models for intracranial hemorrhage detection: strongly annotated computed tomography images and model ensembles

• DOI: 10.3389/fneur.2023.1321964
• Publication Date: 2023-12-29T04:57:26Z
• Authors: Dong-Wan Kang; Gi-Hun Park; Wi-Sun Ryu; Dawid Schellinger...; Museong Kim; Yong Soo Kim; Chan-Young Park; Keon-Joo Lee; Moon-Ku Han; Han-Gil Jeong; Dong-Eog Kim
• Keywords: 0301 basic medicine; Medical Sciences; neuroimaging; Bioinformatics; 610; Diseases; intracranial hemorrhage (ICH); Hematology; deep-learning algorithm; weighted ensemble model; strongly annotated dataset; Biomedical Informatics; 0302 clinical medicine; Oncology; Neurology; Hemic and Lymphatic Diseases; Medical Specialties; Medicine and Health Sciences; Neurology. Diseases of the nervous system; RC346-429

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 algorithm for ICH trained on strongly and weakly annotated datasets, assess whether weighted ensemble model that integrates separate models datasets with different improves performance. used brain CT scans from Radiological Society North America (27,861 scans, 3,528 ICHs) AI-Hub (53,045 7,013 training. DenseNet121, InceptionResNetV2, MobileNetV2, VGG19 were compared independent external test datasets. then developed combining all ICH, subdural (SDH), subarachnoid (SAH), small-lesion cases. The final was four well-known models. After testing, six neurologists reviewed 91 cases difficult AI humans. outperformed when A SDH, SAH, had higher AUC, only. This (AUC [95% C.I.]: Ensemble model, 0.953[0.938-0.965]; 0.852[0.828-0.873]; 0.875[0.852-0.895]; VGG19, 0.796[0.770-0.821]; 0.650[0.620-0.680]; p < 0.0001). In addition, case review showed better understanding management may facilitate use algorithms. propose detection, large-scale, as no can capture aspects complex tasks.