A randomized multi-window 3D deep learning approach for intracranial hemorrhage detection on non-contrast head CT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Intracranial hemorrhage (ICH) is a critical condition requiring swift diagnosis.
• Head computed tomography (CT) is the primary imaging tool for suspected ICH.
• 3D deep learning shows promise for automated ICH detection but faces computational challenges.

Purpose of the Study:

• To introduce Lite3DNet, a lightweight 3D deep learning architecture for automated ICH detection.
• To address the limitations of high computational cost and inference latency in existing models.
• To evaluate Lite3DNet's performance on large-scale public datasets.

Main Methods:

• Developed Lite3DNet, a lightweight architecture utilizing randomized multi-window 3D input.
• Utilized the RSNA ICH Detection dataset, converting DICOM series to 3D volumes.
• Employed 5-fold cross-validation for training and evaluated on internal and external test sets (CQ500, MosMed).

Main Results:

• Lite3DNet demonstrated high discriminative performance across multiple datasets.
• Achieved an AUC of 0.986 on the RSNA internal test set, 0.956 on CQ500, and 0.967 on MosMed.
• The model features 1.189 million parameters and 19.84 ms GPU inference latency.

Conclusions:

• Lite3DNet provides a computationally efficient solution for automated ICH detection.
• The model's low latency and high accuracy support its use in clinical triage.
• This approach enhances the real-world deployability of deep learning for emergency diagnostics.