Computed Tomography (CT) imaging is a prevalent diagnostic modality in healthcare, used to detect abnormalities in the human body. Despite its efficacy, CT scans expose patients to electromagnetic radiation, which can pose health hazards. In response, Low-Dose CT (LDCT) imaging was developed to reduce radiation exposure. However, LDCT degraded image quality, increasing noise and artifacts and decreasing contrast in feature edges. The task of removing noise and artifacts from LDCT images while retaining vital information remains challenging. Traditional noise reduction methods, regardless of their widespread use, have limitations: they are computationally expensive, cause image blurring, and need complex sinogram data. Deep learning-based image-denoising techniques have emerged as a possible alternative to conventional techniques. These techniques have the potential to minimize noise and artifacts from LDCT images while maintaining vital information. This paper presents a novel custom loss function designed to address the noise and artifacts by integrating three key components: Mean Squared Error (MSE), Perceptual loss, and Smoothness regularization (MPS) in a U-Net framework i.e. called MPS-UNet, where MSE quantifies pixel-level differences between original and predicted CT images, ensuring accurate image reconstruction. Conversely, perceptual loss assesses high-level feature distinctions, prioritizing perceptual likeness over pixel accuracy. Meanwhile, smoothness regularization enhances visual quality by penalizing high-frequency variations, fostering smooth transitions between pixels. The proposed model MPSUNet has been tested on the abdominal ‘2016 Low-dose CT AAPM Grand Challenge’ dataset and compared with state-of-the-art algorithms. The results demonstrate superior effectiveness in qualitative and quantitative evaluations, showcasing noteworthy outcomes compared to existing methodologies.