Radiological features of brain hemorrhage through automated segmentation from computed tomography in stroke and traumatic brain injury

• DOI: 10.3389/fneur.2023.1244672
• Publication Date: 2023-09-28T12:56:18Z
• Authors: Bradley J. MacIntosh; Qinghui Liu; Till Schellhorn; Mona K. Beyer; Inge Rasmus Groote; Pål C. Morberg; Joshua M. Poulin; Maiken N. Selseth; Ragnhild C. Bakke; Aina Naqvi; Amir Hillal; Teresa Ullberg; Johan Wassélius; Ole M. Rønning; Per Selnes; Espen S. Kristoff...; Kyrre Eeg Emblem; Karoline Skogen; Else C. Sandset; Atle Bjørnerud
• Keywords: Stroke; Brain hemorrhage

Introduction Radiological assessment is necessary to diagnose spontaneous intracerebral hemorrhage (ICH) and traumatic brain injury intracranial (TBI-bleed). Artificial intelligence (AI) deep learning tools provide a means for decision support. This study evaluates the segmentations produced from three-dimensional AI model that was developed using non-contrast computed tomography (CT) imaging data external current study. Methods Non-contrast CT 1263 patients were accessed across seven sources (referred as sites) in Norway Sweden. Patients included based on ICH, TBI-bleed, or mild TBI diagnosis. Initial images available all participants. Hemorrhage location frequency maps generated. The number of estimated haematoma clusters correlated with total volume. Ground truth expert annotations one ICH site; hence, comparison made volumes. Segmentation volume estimates used receiver operator characteristics (ROC) analysis samples (i.e., bleed detected) then specifically site few TBI-bleed cases. Results showed spatial patterns lesions consistent presentations. There positive correlation between each (correlation range: 0.45–0.74; p -value < 0.01) evidence between-site differences. Relative hand-drawn site, VIOLA-AI segmentation mask achieved median Dice Similarity Coefficient 0.82 (interquartile 0.78 0.83), resulting small overestimate by 0.47 mL 0.04 1.75 mL). detection ROC whole sample gave high area-under-the-curve (AUC) 0.92 (with sensitivity specificity 83.28% 95.41%); however, when considering only head an AUC 0.70. Discussion An open-source tool visualize locations multiple revealed quantitative automated estimate per-participant cluster count. results moderate-to-high. had promising might be useful various types hemorrhage.