Abstract:
Computed Tomography Perfusion (CTP) is a widely used imaging technique for evaluating acute stroke. Compared to other imaging modalities, such as non-contrast computed to...Show MoreMetadata
Abstract:
Computed Tomography Perfusion (CTP) is a widely used imaging technique for evaluating acute stroke. Compared to other imaging modalities, such as non-contrast computed tomography and X-ray, CTP involves a higher radiation exposure because of repeated volume imaging. Thus, reducing the radiation dose is an important task. However, reducing the radiation dose increases noise, which hinders the image quality. This work proposes a machine learning method using a denoising autoencoder (DAE) to remove the noise resulting from dose reduction. The study was conducted using CTP images from the PRove-IT dataset with 46 acute ischemic stroke patients. Low-dose images were simulated by introducing two types of noise, Gaussian noise and Poisson noise, for different reduction levels. These simulated noise images, along with the original images, were used to train the DAE. Additionally, the original perfusion parameter maps were compared with one from different levels of dose reduction. The findings indicate that by reducing the dose to 80%, the original and denoised images were almost identical, with an average Structural Similarity Index (SSIM) of 0.959 and 0.925 for the Gaussian and Poisson noise, respectively. Although the CTP maps had a slightly lower SSIM, they can still determine the core infarction and penumbra, with average SSIM values of four maps of 0.942 and 0.910 for the Gaussian and Poisson groups, respectively. This study demonstrates the potential to decrease the radiation dose of CTP scans by up to 80% with limited compromise on the quality of the resulting images.
Date of Conference: 27-30 May 2024
Date Added to IEEE Xplore: 22 August 2024
ISBN Information: