Deep Learning Framework Predicts Hematoma Expansion in Intracerebral Hemorrhage

Overview

A modular deep learning pipeline using non-contrast CT scans was developed to predict hematoma expansion (HE) in spontaneous intracerebral hemorrhage (sICH). The framework integrates automated segmentation, synthetic data augmentation, and Vision Transformer classification, achieving robust performance across multicenter datasets.

Background

Hematoma expansion occurs in 20-30% of sICH patients and significantly worsens outcomes, increasing mortality and disability risk. Early identification of patients at risk is critical due to a narrow therapeutic window for intervention, typically within 3-6 hours of symptom onset. Non-contrast CT is the frontline imaging modality but current predictors have limitations including variability and limited availability in acute settings. Deep learning offers promise but is challenged by limited data and class imbalance, which synthetic data augmentation may help overcome.

Data Highlights

Key Findings

• The U-Mamba model outperformed other 3D segmentation architectures with a Dice score of 0.937 and IoU of 0.934, demonstrating superior accuracy and robustness in hematoma segmentation.
• Synthetic data augmentation using the Diffusion-UKAN model balanced class distributions, improving classifier training.
• The Vision Transformer classifier trained on a semibalanced dataset (HE:NHE = 1:2) achieved a training AUC of 0.815.
• External validation showed strong generalizability with AUCs of 0.793 and 0.781 on two independent cohorts.
• The modular pipeline enables rapid, automated HE risk stratification using widely available non-contrast CT scans.

Clinical Implications

This deep learning framework facilitates early and accurate identification of sICH patients at high risk for hematoma expansion, enabling timely intervention within the critical therapeutic window. Its automation and reliance on non-contrast CT scans support integration into acute care workflows, potentially improving clinical decision-making and patient outcomes.

Conclusion

The proposed modular deep learning approach combining automated segmentation, synthetic augmentation, and Vision Transformer classification provides a promising tool for rapid and reliable hematoma expansion prediction in sICH. Its robust performance across multicenter datasets underscores its potential for clinical translation.

References

1. Study Authors/2024 -- A Comprehensive Deep Learning Framework for Predicting Hematoma Expansion in Spontaneous Intracerebral Hemorrhage Utilizing Non-Contrast CT Scans