Deep Learning Models for Low Dose CT Simulation

Finding the optimal balance between reducing radiation risk for patients and preserving image quality in CT imaging remains a challenging task. To explore the impact of lower radiation doses on CT image quality, the simulation of low dose CT images is highly demanded. Traditional methods often rely on generating Gaussian/Poisson noise on raw data using mathematical models for specific scanners, which requires access to many medical resources. Inspired from image denoising models, we propose two deep learning models (ResNet and U-Net) for low dose CT images simulation, among which the U-Net model demonstrated better performance. Moreover, the trained U-Net model proves its applicability to untrained CT images from various scanners, showcasing its potential as a versatile and generic low dose CT simulation tool. By bypassing the need for complex phantom experiments and mathematical modeling, deep learning for low dose CT simulation emerges as a promising and powerful approach, akin to solving the inverse problem of well-established image denoising techniques.