Evaluation of Low-dose Computed Tomography Images Reconstructed Using Artificial Intelligence-based Adaptive Filtering for Denoising: A Comparison with Computed Tomography Reconstructed with Iterative Reconstruction Algorithm.

PURPOSE

Awareness of radiation-induced risk led to the development of various dose optimization techniques in iterative reconstruction (IR) algorithms and deep learning algorithms to improve low-dose image quality. PixelShine (PS) by AlgoMedica Inc., USA, is a vendor-neutral deep learning denoising tool for low-dose studies, and this study analyzed its images.

AIM

The aim of this study was to assess the diagnostic value of PS-reconstructed images obtained at various low doses (LDs).

MATERIALS AND METHODS

A retrospective study qualitatively and quantitatively evaluated the low-dose PS-reconstructed images by comparing them with other reconstruction methods and standard dose (SD) images. A total of 85 cases were evaluated, of which 32 cases were scanned on a scanner with filtered back projection (FBP) reconstruction with LD scans performed at 70%-50% of SD. The remaining 53 cases were performed on the scanner with IR, 35 of them had LD scan at 50% of SD and 18 cases had LD scan at 33% of SD.

RESULTS

Qualitative image analysis - The quality of low-dose images with PS and IR was almost equivalent in terms of noise magnitude and texture at 50% dose, and PS images were slightly better at 33% dose reduction. Quantitative image analysis - Low-dose PS-reconstructed images and low-dose iterative reconstructed images had similar contrast-to-noise ratio at 50% dose reduction; however, at 33% of the SD, PS-reconstructed images outperformed. The SD FBP images were equivalent to LD PS-reconstructed images (50% dose reduction).

CONCLUSIONS

Artificial intelligence-based denoising algorithms produce similar images as IR at 50% dose reduction and outperform it at 33% of the SD.