Low-dose CT reconstruction by self-supervised learning in the projection domain.

OBJECTIVE: To address the critical issue of compromised image quality and diagnostic accuracy in low-dose computed tomography (LDCT) due to increased noise and artifacts, we aim to develop a self-supervised learning model, Noise2Projection, that enhances LDCT image quality without the requirement for paired CT images. This objective is crucial for mitigating the adverse effects of excessive X-ray radiation exposure on patients while ensuring the reliability of clinical diagnosis. METHODS: We devised a self-supervised learning approach that leverages the inherent correlation between raw noisy CT projection images to reduce noise and artifacts. The Noise2Projection model was trained and validated using a dataset of clinical LDCT scans, employing a novel algorithm that does not rely on paired images, thereby overcoming a significant limitation in clinical practice. RESULTS: Both quantitative and qualitative assessments of the model's output revealed substantial improvements in LDCT image quality. The model effectively reduced noise levels and eliminated artifacts, resulting in images that were more suitable for clinical interpretation. The performance metrics indicated a clear enhancement in diagnostic image quality without the need for additional radiation or paired training data. CONCLUSION: The Noise2Projection model represents a significant advancement in the field of LDCT imaging, providing a self-supervised solution that significantly improves image quality and reduces artifacts. This innovation not only contributes to the reduction of patient radiation exposure but also ensures the delivery of high-quality images essential for accurate clinical diagnosis.