Comparative evaluation of deep learning-based and conventional reconstruction techniques for image quality enhancement in low-dose chest computed tomography.

Background: With the increasing use of computed tomography (CT) in clinical practice, greater attention is being paid to the radiation dose. There are numerous methods available for reducing the CT radiation dose, but enhancement via a reconstruction algorithm is the most effective method. This study aimed to evaluate the performance of a proposed deep learning (DL) algorithm for low-dose chest CT image reconstruction among patients with pulmonary diseases and to compare it with several mainstream iterative reconstruction (IR) techniques. Methods: A retrospective analysis was performed on 90 patients who attended hospital for chest CT physical examination. Among them, 45 persons for physical check-up underwent low-dose chest CT, and 45 were outpatients who underwent conventional chest CT. Raw data from low-dose CT were used for reconstruction via IR and filtered back projection (FBP). The images reconstructed using FBP in standard-dose CT served as the gold standard. The DL algorithm took the low-dose FBP as input and output denoised images. Image quality was assessed using the signal-to-noise ratio (SNR), absolute noise, the contrast-to-noise ratio (CNR), the mean squared error (MSE), and image structural similarity (SSIM). Moreover, two experienced radiologists subjectively and blindly evaluated the results using a five-point scale. Results: Compared with IR reconstructions, the DL method had significantly higher SNR and CNR and lower MSE and absolute noise, while the SSIM and subjective quality scores remained statistically similar. No significant difference in the 5-point scale was found between the DL-enhanced images and FBP-reconstructed standard-dose images. In addition, DL method reduced the radiation dose up to 91% as compared to standard-dose scans. Conclusions: The proposed DL method provided better image quality than did the mainstream IR methods in low-dose chest CT images. This novel DL technique could be effective and practicable for significantly reducing radiation dose while ensuring the quality of clinical diagnostic images in routine radiology work.