Deep Learning-Based CT Image Quality Enhancement: Challenges and Future Directions

CT is an important diagnostic modality in clinical medicine, which converts the absorption of X-rays by the human body into internal morphology of anatomical structures such as bones, muscles, and tissues, by exploiting the heterogeneity and density differences of human tissues. It is widely used for early diagnosis and postoperative intervention. However, the quality of CT is affected by artifacts due to severe scatter noise and truncation projection, such as shadows, calcifications, and metal, making it unsuitable for accurate dose calculation and affecting imaging quality and subsequent diagnosis. Although existing deep learning-based methods have achieved promising success in CT image quality enhancement, most of these methods follow the idea of classical algorithms for this task, using dual-domain analysis of image and signal domains to remove artifacts. However, this framework is difficult to cope with the artifact patterns caused by complex metal implants and can introduce new noise in the mapping process.