Learning non-local perfusion textures for high-quality computed tomography perfusion imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Computed tomography perfusion (CTP) imaging is vital for acute stroke assessment due to its accessibility and speed.
• Standard CTP protocols involve significant radiation doses, raising concerns about potential long-term cancer risks.
• Developing low-dose CTP techniques is crucial to mitigate radiation exposure while maintaining diagnostic accuracy.

Purpose of the Study:

• To introduce a novel deep learning model, the non-local perfusion texture learning network (NPTN).
• To enable high-quality CTP imaging from low-dose scans.
• To improve the accuracy of cerebral blood flow (CBF) estimation in low-dose CTP.

Main Methods:

• Developed a non-local perfusion texture learning network (NPTN) utilizing deep learning.
• Incorporated non-local self-similarities from adjacent frames to construct texture vectors.
• Employed a convolutional neural network with residual learning and batch normalization for image reconstruction.

Main Results:

• NPTN demonstrated superior performance over competing methods in low-dose CTP reconstruction.
• Achieved approximately 3.0 dB higher average PSNR, 1.4% increase in average SSIM, and 4.8% decrease in average RMSE.
• Improved CBF estimation with a 3.4% increase in average SSIM and a 61.1% decrease in average RMSE.

Conclusions:

• The NPTN method successfully generates high-quality CTP images and accurate CBF maps from low-dose data.
• The model effectively characterizes structural details and contrast variations, outperforming existing methods.
• NPTN offers a promising solution for reducing radiation dose in CTP imaging for stroke assessment.