Convolutional Neural Networks Predict Early Hematoma Enlargement In Intracerebral Hemorrhage Patients Based On Non-contrast Computed Tomography Images

Objective: Spontaneous intracerebral hemorrhage(sICH)is one of the most dangerous types of stroke,with high morbidity,high mortality,high disability and high recurrence rate.There is evidence that hematoma enlargement(HE)is a preventable and independent predictor of early neurological deterioration and poor long-term outcome,and few interventions ultimately improve clinical outcomes after ICH.Therefore,limiting HE is an important goal in the acute care of ICH and in the prevention of adverse outcomes.Non-contrast computed tomography(NCCT)has the characteristics of rapid,easy to obtain,and high sensitivity in identifying acute ICH so it is considered the gold standard for the diagnosis of ICH in the emergency department.Artificial intelligence,especially the convolutional neural network(CNN)in deep learning algorithms,can achieve advanced performance in the field of medical image analysis with high accuracy and sensitivity.Among them,YOLOv3 based on CNN is currently one of the most powerful,fast and advanced deep learning object detection algorithms.YOLO algorithm is a series of algorithms for stage object detection.Its network structure can be directly used to extract target features to achieve accurate detection and classification,so YOLO algorithm is more simple and efficient than the traditional CNN model.The purpose of this study is to use YOLOv3 network to establish a prediction model based on NCCT images of patients with ICH to predict the risk of early HE in patients with ICH,in order to quickly screen the high-risk groups with HE by directly inputting NCCT images into the deep learning prediction model,so as to achieve the goal of providing basis for patients to develop personalized clinical treatment plans and guidelines.Methods: A total of 779 patients with sICH admitted to Third-class hospital from January 2010 to December 2020 were retrospectively collected as the training set(70%)and test set(30%)data of the deep learning prediction model.A total of 332 patients with sICH admitted to our hospital from January 2021 to June 2022 were prospectively collected as an external validation set.The imaging data,clinical data and clinical indicators of the patients were collected.Early hematoma expansion was defined as the increase of hematoma volume by>33% or>6 ml after the first scan of the head CT within 24 hours after the onset of the disease.The YOLOv3 integrated framework is used to identify the HE images of ICH,establish the prediction model of HE,and use the mean average precision(Map)as the evaluation index of the prediction performance of the model.In addition,based on univariate analysis to screen out clinical indicators affecting the occurrence of HE in patients with ICH,a nomogram prediction model was constructed to further evaluate the predictive value of clinical indicators for the occurrence of HE in patients with ICH.Results: A total of 1111 patients with sICH were enrolled.The mean age of the patients was 61.21±13.46,and 739(66.5%)were male.Of these 779 patients included in the training set,175(22.5%)developed HE with ICH;332 patients were included in the validation set,51(15.4%)developed HE with ICH.The m AP value of YOLOv3 deep learning prediction model established based on NCCT images to predict HE of ICH reached 0.912.The AUC value of the visual nomogram prediction model based on the clinical indicators was 0.715(0.673-0.758)in the training set and 0.644(0.554-0.734)in the validation set.Conclusions: The YOLOv3 network model based on NCCT images can effectively predict HE in patients with ICH,while the prediction efficacy of the visual nomogram model based on clinical data alone was general.