Noncontrast CT-Based Deep Learning for Predicting Hematoma Expansion Risk in Patients with Spontaneous Intracerebral Hemorrhage

Overview

Hematoma expansion is an independent predictor of poor prognosis and early neurological deterioration in patients with spontaneous intracerebral hemorrhage. Early identification of high-risk patients and timely targeted medical interventions may provide a crucial opportunity to limit hematoma growth and improve neurological outcomes. This study aims to develop an end-to-end deep learning model based on noncontrast computed tomography images to predict the risk of hematoma expansion in patients with spontaneous intracerebral hemorrhage. This model could serve as a valuable risk stratification tool for patients with hematoma expansion, facilitating targeted treatment and providing clinicians with streamlined decision-making support in emergency situations.

This project is planned to be implemented in four steps: 1\. Data Collection 1. Selection of Study Subjects: Clinical and imaging data of patients with spontaneous intracerebral hemorrhage were retrospectively collected from multiple centers, including 500 cases in the hematoma expansion group and 1500 cases in the non-expansion group, totaling 2000 cases. Hematoma expansion (rHE) was defined as an absolute increase in ICH volume of ≥6 mL or a relative increase of ≥33%. 2. Collection of Clinical Data: Includes patient age, gender, history of coronary heart disease, smoking, alcohol, hypertension, admission systolic and diastolic blood pressures, among others. 3. CT Image Acquisition: Admission and follow-up CT images were obtained using spiral CT scanning with a slice thickness and interslice spacing of 5 mm. 2\. Segmentation of Hematoma Based on Non-contrast CT Images Two radiologists independently segmented the volume of interest of the entire brain hematoma lesion using ITK-SNAP software, manually outlining the lesion on each CT slice while avoiding the surrounding edema and normal brain tissue. 3\. Establishment of Automatic Hematoma Segmentation Model 1. Data Acquisition and Preprocessing: All images were obtained through the PACS system and stored in DICOM format. Standardized preprocessing steps were applied, including image resampling, window width, and window level adjustments to accommodate parameter differences across different CT scanners. 2. Selection of Automatic Segmentation Model: Suitable deep learning architectures for segmentation were explored and selected, including encoder-decoder structures such as nnU-Net, UNETR, and nnFormer. The optimal image segmentation model was chosen to achieve precise segmentation of brain hematoma regions. 3. Model Training and Evaluation: The model was trained using supervised learning, with manually segmented masks from the annotated dataset serving as ground truth labels. Model performance was evaluated on validation and independent external test sets using metrics such as Dice coefficient, Intersection over Union (IoU), precision, and recall. 4\. Establishment of Automatic Classification Model for Hematoma Expansion 1. Construction of the Automatic Classification Model: Based on the segmentation masks extracted by the automatic segmentation model, a deep learning classification model was developed to predict hematoma expansion. Various 2D and 3D classification neural networks, including 2D-ResNet-101, 2D-ViT, 3D-ResNet-101, and 3D-ViT, were developed. Using the 3D masks generated by automatic segmentation, the largest 2D rectangular region of interest and the smallest 3D bounding box of the brain hematoma were cropped from the original CT images, and these cropped regions were input into the corresponding deep learning classification models to achieve precise prediction of hematoma expansion. 2. Visualization of the Automatic Classification Model: To visually verify the decision-making process of the deep learning model, Gradient-weighted Class Activation Mapping (Grad-CAM) technology was used to generate 2D attention maps, visually displaying the key hematoma regions identified by the model for classification. 3. Model Training and Evaluation: During model evaluation, the performance of the model was tested using an independent external test set, with comprehensive evaluation metrics including accuracy, sensitivity, specificity, F1 score, ROC curve, and AUC value. This process aimed to validate the model\'s generalizability and robustness across multicenter data, ensuring its reliability and effectiveness in actual clinical applications.