Multiphase Computed Tomographic Angiography with Bone Subtraction Using 3D Multichannel Convolution Neural Networks^.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Multiphase computed tomographic angiography (CTA) is crucial for assessing cerebral collateral circulation in acute ischemic stroke.
• Selecting patients for recanalization therapy relies on accurate evaluation of collateral status.
• Current imaging triage can be time-consuming, necessitating faster and more consistent decision-making.

Purpose of the Study:

• To develop and evaluate a bone subtraction technique for multiphase CTA to improve visualization of cerebral collateral circulation.
• To enable faster and more consistent clinical decisions in the imaging triage of acute stroke patients.
• To utilize deep learning for automated bone mask extraction from contrast-enhanced CTA scans.

Main Methods:

• Forty multiphase brain CTA datasets were analyzed.
• Two bone subtraction methods were compared: a reference method using pre-contrast scans and a tested method using contrast-enhanced scans.
• Two 3D multichannel convolutional neural network (CNN) architectures (U-net and atrous CNN) were trained and tested for bone mask segmentation using a supervised deep learning approach.

Main Results:

• The tested U-net CNN achieved a mean intersection over union (IoU) of 90.0% +/- 2.2.
• The patch-based 3D multichannel atrous CNN achieved a superior mean IoU of 93.9% +/- 1.2.
• The bone subtraction technique successfully visualized CTA volumetric datasets, creating full brain angiogram-like images.

Conclusions:

• Deep learning-based bone subtraction is a reliable technique for enhancing multiphase CTA visualization.
• This method assists clinicians in the emergency department by providing clear images for evaluating acute ischemic stroke patients.
• The technique supports faster and more consistent decision-making in acute stroke imaging triage.