Effectiveness of deep learning-based denoising on image quality and diagnostic performance of low-dose abdominal CT...

To evaluate the effect of deep learning-based denoising on image quality and diagnostic performance of low-dose abdominal CT in diagnosing acute appendicitis, and to determine whether filtered back projection (FBP) or iterative reconstruction (IR) provides more suitable input images for denoising.

This single-center retrospective study included 100 patients (40 men, 60 women; median age, 31 years) who underwent low-dose abdominal CT (median effective dose, 1.34 mSv) for suspected acute appendicitis between May 2023 and December 2024. Images were reconstructed using FBP and IR, then processed with vendor-neutral, deep learning-based denoising software, generating four image series per patient (FBP, denoised FBP, IR, and denoised IR). Two radiologists independently analyzed image noise, sharpness, artifacts, and overall quality using 4-point scales and assessed the presence of acute appendicitis. Quantitative noise and signal-to-noise ratio (SNR) were measured from circular regions of interest in reference tissues. Diagnostic performance and image quality metrics were compared across series using appropriate statistical methods.

Noise and SNR differed significantly among series (both p < 0.001). Quantitative noise was lowest for denoised IR, followed by denoised FBP, IR, and FBP, while SNR showed the opposite trend. Qualitative noise and sharpness scores were significantly higher for denoised series. Overall image quality was highest for denoised FBP, followed by denoised IR, IR, and FBP. Sensitivity and specificity for diagnosing acute appendicitis showed no significant inter-series differences; however, specificities were numerically higher for denoised series.

Deep learning-based denoising improved quantitative and qualitative image quality in low-dose abdominal CT for acute appendicitis, although no significant differences in diagnostic performance were observed. Denoised FBP showed the most favorable subjective image quality.