基于CT图像的高精度血肿检测与脑血肿扩大预测模型研究

A High-precision Prediction Model Based on CT Image for Hematoma Detection and Cerebral Hematoma Enlargement

目的:开发一种可无缝整合脑出血(CH)患者的血肿扩大(HE)成像数据和临床信息的自动化模型,并评估其预测价值。方法:回顾性收集2016年9月至2020年10月在深圳市第二人民医院及深圳市罗湖区人民医院行颅脑计算机断层扫描(CT)的患者及其临床基线资料。模型训练分为两部分,第一部分是开发出血类型识别模型,纳入了深圳市第二人民医院2836例病例资料;第二部分为开发HE预测模型,纳入了530例CH患者的资料,包括228例HE阳性患者和302例HE阴性患者。外部临床验证队列由深圳市第二人民医院及深圳市罗湖区人民医院收集的共9677例患者组成,与第一队列无重叠。结果:对于出血类型分类模型,2分类模型(是否有CH)的准确度、灵敏度与特异度都达到了100.00%。对于HE病例的预测,基于图像和临床特征训练的深度学习模型在内部测试集和临床验证队列中的受试者工作特征曲线下面积(AUROC)分别为0.865和0.862。结论:深度学习模型对HE的预测快速准确,表明了在临床环境中改进诊断和治疗计划程序的潜力。

Objective Develop an automated model that seamlessly integrates hematoma enlargement(HE)imaging data and clinical information in patients with cerebral hemorrhage(CH)and evaluate its predictive value.Methods Patients’clinical and CT data,who underwent craniocerebral computed tomography(CT)in Shenzhen Second People's Hospital or Shenzhen Luohu District People's Hospital,were retrospectively collected between September 2016 and October 2020.The model training was divided into two parts.The first part was to develop the hemorrhagic type recognition model,which included 2836 cases of Shenzhen Second People's Hospital.The second part was to develop the HE prediction model,which included data from 530 patients with CH,including 228 patients who were positive for HE and 302 patients who were negative for HE.The external clinical validation cohort consisted of a total of 9677 patients collected from two hospitals(Shenzhen Second People's Hospital and Shenzhen Luohu District People's Hospital),which did not overlap with the first cohort.Results For the bleeding type classification model,the accuracy,sensitivity,and specificity of the 2-class model(with or without CH)all reached 100.00%.For the prediction of HE cases,the area under the receiver operation characteristic curve(AUROC)of the deep learning model based on image and clinical feature training in the internal test set and external validation cohort was 0.865 and 0.862,respectively.Conclusion The deep learning model for predicting HE can quickly and accurately predict the occurrence of HE,indicating the potential to improve diagnostic and treatment planning procedures in clinical settings.