Comparing manual and automated methods for extracting perihematomal features in predicting hemorrhage expansion in deep intracerebral hemorrhage

This study aimed to compare the efficacy of radiomics models based on automated perihematomal tissue expansion versus manual perihematomal edema (PHE) delineation for predicting hematoma expansion (HE) within 72 h in patients with deep intracerebral hemorrhage (ICH). A retrospective collection of patients with ICH from three medical centers was divided into training ( n = 219) and validation ( n = 55) sets using an 8:2 ratio while collecting an independent external test ( n = 69) set. Based on CT within 24 h after ICH, employing machine learning, eight models were constructed by integrating hematoma features with automatically expanded perihematomal regions (ranging from 4 to 18 mm in 2-mm intervals) to predict HE. Then, each model was individually compared with the manually delineated Intra + PHE radiomics model. The models were evaluated using the area under the curve (AUC) and decision curve analysis. The Shapley additive explanation method was applied to highlight key features. In the test set, the five Intra+AutoPeri radiomics models demonstrated a higher AUC than that of the Intra + PHE model. Specifically, the Intra+AutoPeri-14 mm model achieved the highest AUC (0.899) in the test set, significantly outperforming the Intra + PHE model (AUC = 0.764, p = 0.039). The automated Intra+AutoPeri model, which demonstrates a promising trend of superior predictive performance over the manually-derived Intra + PHE model, holds significant promise for stratifying the risk of HE in deep ICH, providing valuable insights for early, individualized patient management.