Texture-Based Preprocessing Framework with nnU-Net Model for Accurate Intracranial Artery Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate segmentation of intracranial arteries from DSA is crucial for diagnosing neurovascular conditions and planning treatments.
• Current deep learning models show promise but are limited by preprocessing techniques.
• Enhancing preprocessing is key to improving the accuracy of vascular extraction.

Purpose of the Study:

• To develop and evaluate a texture-based contrast enhancement preprocessing framework integrated with nnU-Net for improved intracranial artery segmentation.
• To assess the impact of the proposed method on segmentation accuracy and topological representation in DSA images.

Main Methods:

• A texture-based contrast enhancement framework was developed, fusing local contrast, entropy, and brightness threshold maps into a combined feature mask.
• This feature mask was used as input for the nnU-Net deep learning model for segmentation.
• Performance was evaluated on the DIAS dataset using metrics like Dice Similarity Coefficient (DICE) and Intersection over Union (IoU).

Main Results:

• The proposed method achieved a DICE of 0.83 ± 0.20 and IoU of 0.72 ± 0.14, outperforming CLAHE and baseline methods.
• Significant improvements in vessel connectivity (VC) and topological accuracy were observed, with VC dropping by over 65% relative to unprocessed images.
• The texture-based preprocessing demonstrated robustness and noise tolerance compared to existing methods.

Conclusions:

• Integrating texture-based preprocessing with nnU-Net substantially enhances intracranial artery segmentation from DSA.
• The method provides robust, noise-tolerant, and clinically interpretable results, advancing neurovascular diagnostic capabilities.
• This approach offers a significant improvement over traditional methods for vascular segmentation in medical imaging.