An Ai-Based Convolutional Auto-Encoder Framework for Low-Dose CT Image Denoising and Quality Enhancement

Abstract

Computed tomography (CT) is a vital medical imaging mode, and radiation dose reduction can result in a higher noise level and worse image quality, which can restrict its use in diagnostics. To overcome this difficulty, this paper will present an AI-based convolutional auto-encoder model to resolve the low dose CT image denoising and quality improvement of paired Quarter Dose CT data and Full Dose CT data. Original DICOM images were utilized so as to have a transparent and reproducible preprocessing pipeline. The technique incorporated metadata-based slice correspondence, Hounsfield Unit transformation, intensity clipping, normalization, patient separate information division, and patch-based supervised education. An untrained basic convolutional auto-encoder was used and trained to produce Full Dose-like images by using Quarter Dose as inputs. The test set that was held out and evaluated experimentally showed that the convergence was stable, and the denoising performance was strong, with a mean MSE of 0.000293, RMSE of 0.015310, MAE of 0.011648, PSNR of 37.3363 dB, and SSIM of 0.8669 Subgroup analysis also revealed that the model worked better with 3 mm slices compared to 1 mm slices and better with the B30 reconstruction kernel compared to D45 kernel with the optimal results being in the 3mm-B30 setting. These results suggest that the suggested framework is useful to reduce the low-dose CT noise and maintain the structural similarity simultaneously, thus offering a feasible and interpretable solution to supervised image restoration of low-dose CT images.