Abstract
OBJECTIVE: This study evaluates the utility of artificial intelligence (AI) for automated segmentation of intracranial hematomas and surrounding oedema in non-contrast computed tomography (CT) images. Additionally, it aims to extract imaging features for developing machine learning models to predict hematoma expansion in acute spontaneous intracerebral hemorrhage (sICH). METHODS: Data from 183 patients with acute spontaneous hemorrhage, treated at Lianyungang Hospital Affiliated to Xuzhou Medical University between January 2020 and December 2023, were retrospectively analyzed. Patients were divided into training (n = 128) and testing (n = 55) sets in a 7:3 ratio. CT images were segmented using United Imaging uAI software and both imaging features and clinical characteristics were extracted. Independent risk factors were identified through univariate analysis and least absolute shrinkage and selection operator (LASSO) regression. Machine learning algorithms were applied to construct predictive models for hematoma expansion. Model performance was evaluated using receiver operating characteristic (ROC) curves and the area under the curve (AUC). RESULTS: Eight feature parameters were extracted from the CT images. The comprehensive model achieved an AUC of 0.9027, with a sensitivity of 0.8235 and specificity of 0.8831. A simplified model utilizing four imaging features yielded an AUC of 0.8897, with a sensitivity of 0.7451 and specificity of 0.9221, slightly underperforming compared to the comprehensive model. Incorporating the subjective 'swirl sign', identified as the most significant feature in univariate analysis, into the simplified model enhanced its performance. This optimized model achieved an AUC of 0.9524, with a sensitivity of 0.9412 and specificity of 0.9091, surpassing both the comprehensive and simplified models. CONCLUSION: The optimized model, based on CT imaging features of hematomas and surrounding oedema, offers a practical and reliable tool for predicting hematoma expansion in sICH. Its robust performance supports its utility in emergency settings to guide clinical decision-making effectively.