Sparse Representation-Based LDCT Image Quality Assessment Using the JND Model

Abstract

In clinical environments, the quality of low-dose computerized tomography (LDCT) images is inevitably degraded by artifacts and noise. Accurate and clinically relevant image quality assessment (IQA) for LDCT images is crucial to provide appropriate medical care. However, due to the absence of reference images and inadequate understanding of the human visual system (HVS), there is currently a lack of effective IQA methods. To address this problem, this paper proposes a no-reference IQA method that emulates the perceptual characteristics of the HVS. The method is based on the use of sparse representation in conjunction with a just noticeable distortion (JND) model. Specifically, for each tested image, a patch-level JND map is first calculated to indicate the noticeable level of distortion in different regions. Subsequently, noticeable patches of the image are encoded via sparse representation, followed by max pooling to obtain sparse features. Finally, these sparse features, along with the sorted JND values, are combined as overall features into a regression model to predict an objective quality score. By combining two types of features, our method can measure the quality from the perspectives of both local spatial structures and visual distortion sensitivity. Our method is evaluated on the LDCTIQAC2023 database, and the experimental results demonstrate the effectiveness of our method, which correlates reasonably well with the radiologists’ subjective scores.

Keywords

• CT images
• image quality assessment
• sparse representation
• just noticeable distortion
• human visual system