A Deep Learning Approach for Automated Bone Removal from Computed Tomography Angiography of the Brain

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Neurovascular Imaging

Background:

• Advanced visualization techniques like maximum intensity projection (MIP) and volume rendering (VR) are crucial for brain CT angiography (CTA).
• Osseous interference in CTA images limits the evaluation of neurovascular anatomy, especially at the skull base.
• Existing bone removal methods for CTA often lack accuracy and scope.

Purpose of the Study:

• To develop and validate a novel deep convolutional neural network-based tool for automated bone removal from brain CTA.
• To enhance the visualization of neurovascular structures by eliminating interfering osseous anatomy.
• To compare the performance of the new bone removal algorithm against existing methods.

Main Methods:

• A deep convolutional neural network was designed and trained on 72 brain CTAs for bone removal.
• The model was rigorously tested on 15 internal and 17 independent external CTAs.
• Quantitative assessment used Dice overlap with manual segmentations; qualitative assessment evaluated VR visualization of carotid siphons.

Main Results:

• The algorithm achieved high accuracy with average Dice overlap scores of 0.986 (internal) and 0.979 (external).
• Performance surpassed a public bone removal algorithm (Dice 0.947 internal, 0.938 external).
• Superior VR visualization of carotid siphons was achieved in 93% (internal) and 88% (external) of cases, with significant p-values.

Conclusions:

• The proposed brain CTA bone removal algorithm is rapid, accurate, and effective.
• It significantly improves the visualization of vascular anatomy and pathology compared to existing techniques.
• Validation on an independent external dataset confirms the algorithm's generalizability and clinical utility.