Explainable CT-based deep learning model for predicting hematoma expansion including intraventricular hemorrhage...

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Hematoma expansion (HE), including intraventricular hemorrhage (IVH) growth, critically impacts patient outcomes following intracerebral hemorrhage (ICH).
• Accurate prediction of HE is crucial for timely intervention and improved patient management in ICH cases.

Purpose of the Study:

• To develop, validate, and interpret a deep learning model, HENet, for predicting three distinct definitions of hematoma expansion.
• To assess HENet's predictive performance against existing 2D models and physician expertise.

Main Methods:

• A multicenter retrospective study involving 718 ICH patients from three hospitals.
• Utilized CT scans and clinical data to train and validate the HENet model for predicting revised hematoma expansion (RHE) definitions 1 and 2, and conventional HE (CHE).
• Employed the Grad-CAM technique for model interpretability.

Main Results:

• HENet demonstrated high AUC values for predicting RHE1, RHE2, and CHE.
• The model significantly outperformed physician predictions and 2D models in net reclassification index and integrated discrimination index for RHE1 and RHE2.
• Visual insights from Grad-CAM aided in understanding the model's decision-making process.

Conclusions:

• The deep learning model HENet shows high accuracy in predicting hematoma expansion in ICH patients.
• Integration of HENet into clinical practice has the potential to enhance prediction accuracy and improve patient outcomes.
• AI-driven prediction tools can offer valuable support in managing intracerebral hemorrhage.