Hybrid 3D/2D Convolutional Neural Network for Hemorrhage Evaluation on Head CT

Area of Science:

• Medical imaging analysis
• Artificial intelligence in radiology
• Deep learning for neuroimaging

Background:

• Convolutional neural networks (CNNs) are advanced tools for image recognition.
• Accurate detection of brain hemorrhages on noncontrast CT (NCCT) is critical for patient management.

Purpose of the Study:

• To evaluate a CNN optimized for detecting and quantifying intraparenchymal, epidural/subdural, and subarachnoid hemorrhages on NCCT.
• To assess the clinical viability of a deep learning tool for hemorrhage evaluation in emergency department settings.

Main Methods:

• A custom hybrid 3D/2D mask ROI-based CNN was developed and cross-validated on a training cohort of 10,159 NCCT examinations.
• The trained network was prospectively applied to 862 NCCT examinations from the emergency department via an automated inference pipeline.
• Performance metrics included accuracy, AUC, sensitivity, specificity, PPV, NPV, Dice scores, and Pearson correlation.

Main Results:

• The CNN achieved high performance in hemorrhage detection, with an accuracy of 0.970 and AUC of 0.981 on the prospective test set.
• Specific hemorrhage types showed varying Dice scores: intraparenchymal (0.931), epidural/subdural (0.863), and subarachnoid hemorrhage (0.772).
• The tool demonstrated strong sensitivity (0.951) and negative predictive value (0.993) for hemorrhage detection.

Conclusions:

• A customized deep learning tool accurately detects and quantifies various types of intracranial hemorrhages on NCCT.
• The demonstrated high performance on prospective emergency department NCCTs supports the clinical viability of this AI tool.