|
The main challenge for low-dose computed tomography (CT) imaging is the poor noise characteristics of resultant images. Photon starvation leads to high noise in projection data and consequently the reconstructed images suffer from noise and streaks. In order to achieve a lower dose during imaging, it is necessary not only to minimize the mA but also to increase the scan speed, thereby reducing the time of radiation exposure to the patient. The cone-beam CT (CBCT) images obtained by applying all the methods of reducing mA and capturing images quickly, resulting in a radiation dose lower than the existing low-dose technique, are defined as ultra-low dose (ULD) images. In order to improve the quality of ULD images, an approach using deep learning methods was attempted. For this, a pair of ULD images and the corresponding target normal dose (ND) image to be improved should be available. However, due to the difficulty of obtaining ULD images, we propose a method of converting existing normal dose CBCT images into ULD images and using them for deep learning training. The ULD images simulated using the proposed method showed minimal differences compared to the images acquired under actual ULD imaging conditions. For the simulated ULD, there was a 2.02% difference in CT numbers compared to the actual ULD, and a 1.69% difference in noise levels. Training deep learning with simulated ULD data, which closely resembles the actual data with minimal differences, using the same parameters as training with real data, yielded performance similar to training with actual data. Through visual examination and quantitative analysis, it was verified that training deep learning on an extensive dataset of simulated ULD data results in improved performance, contrasting with the challenges associated with obtaining actual data. Utilizing the proposed approach is anticipated to enable the effective application of deep learning in demanding medical domains where obtaining real data presents challenges, ultimately leading to the attainment of desired results.
|
