Prediction of Early Hematoma Expansion in Spontaneous Intracerebral Hemorrhage Patients without Conventional Radiological Signs By Deep Learning Features
10.3969/j.issn.1005-5185.2024.12.004
- VernacularTitle:基于深度学习特征预测无传统影像学征象的自发性脑出血患者早期血肿扩大
- Author:
Wanjun LU
1
;
Jian PENG
;
Mengxuan YUAN
;
Liqing GAO
;
Jieling SHEN
;
Chengtuan SUN
Author Information
1. 扬州大学附属江都人民医院神经内科,江苏 扬州 225200
- Publication Type:Journal Article
- Keywords:
Spontaneous intracerebral hemorrhage;
Hematoma expansion;
Deep learning;
Radiomics;
Forecasting;
Deep residual network
- From:
Chinese Journal of Medical Imaging
2024;32(12):1215-1221
- CountryChina
- Language:Chinese
-
Abstract:
Purpose To explore the value of deep learning feature prediction based on the ResNet50 deep residual network model for predicting early hematoma expansion in spontaneous intracerebral hemorrhage without traditional imaging manifestations. Materials and Methods A retrospective study was performed on 235 patients with spontaneous intracerebral hemorrhage in Jiangdu People's Hospital Affiliated to Yangzhou University from January 2019 and December 2022. These patients had undergone their initial plain cranial CT scan within 6 hours of symptom onset and a subsequent follow-up scan within 24 hours of admission. They were randomly assigned to a training set consisting of 188 cases and a test set of 47 cases,using an 8︰2 ratio. The region of interest (ROI) of hematoma was traced layer by layer on the first plain head CT,and image genomics features were extracted. The maximum two-dimensional cross-sectional ROI of the hematoma 3D-ROI,as well as ROI images at 1 mm and 2 mm above and below the maximum two-dimensional cross-sectional ROI,were then cut and input into the pre-trained ResNet50 model for feature extraction. The image genomics features were then fused with the extracted deep learning features using a least absolute shrinkage and selection operator regression model. A support vector machine classifier was used to construct a prediction model,which was evaluated using receiver operating characteristic curves and decision curve analysis. Results In the training set,the area under curve (AUC) of the deep learning feature model was 0.972,which was higher than that of the image genomics feature model (0.951) and the fused feature model (0.968),but this difference was not statistically significant (P>0.05). In the testing set,the AUCs of the deep learning feature model and the fused feature model were 0.867 and 0.895,respectively,which were significantly higher than that of the image genomics feature model (0.833),with statistically significant differences (Z=-1.794,-2.191,both P<0.05). The AUC of the fused feature model showed an improvement compared to the deep learning feature model,but the difference was not statistically significant (P>0.05). In the test set,decision curve analysis revealed that the fused feature model yielded greater benefits compared to both the deep learning feature model and the radiomic feature model. Conclusion The deep learning feature model based on ResNet50 deep residual network shows better performance in predicting early hematoma expansion than the image genomics feature model,and the fused feature model has a beneficial effect on predicting hematoma expansion. This deep learning approach provides a prediction tool with supervisory capability for clinical decision-making.
