Enhancing Low-Dose CT Image Quality Through Deep Learning: A DoG-Sharpened U-Net Approach With Attention Mechanism

Computed Tomography (CT) scans utilize electromagnetic radiation. However, the excessive exposure of a patient's body during CT acquisition poses potential health risks. Subsequently, the integration of low-dose CT scans has resulted in an escalation of noise, artifacts, and a discernible decline in the overall quality of CT imaging, significantly impacting the diagnostic capabilities of Computer-Aided Diagnosis (CAD) system. Removing these noises and artifacts while preserving critical information is a challenging task. Traditional noise reduction algorithms are expensive, produce blurry results, and rely on challenging sinogram data. As a result, deep learning-based image-denoising approaches have emerged. This study introduces a DoG-UNet+ model which incorporates a novel layer called the “Difference of Gaussians (DoG) Sharpening Layer” into the U-Net architecture. This layer utilizes two distinct convolutional kernels, termed “fat” and “skinny,” aimed to capturing diverse scales of features. To increase the clinical detection precision, an attention mechanism has been added to focus on the critical features in CT images. This model has been compared with the most recent algorithms based on the Peak Signal-to-Noise Ratio (PSNR), Root Mean Squared Error (RMSE), and Structural Similarity Index (SSIM). The outcomes showed superior performance of DoG-UNet+ model, showcasing promising and notable results compared to existing methods.