SR-CNN: Outlier Detection Preprocessing for Low-Dose CT Noise Equalization

Abstract

Low-dose computed tomography (LDCT) mitigates X-ray risks yet yields noisy images. Conventional unified denoising ignores spatially uneven noise, failing to suppress high-noise region interference and limiting performance. The core of this work is to equalize the spatially uneven noise distribution of LDCT images via a dedicated preprocessing strategy, so as to fully activate the denoising potential of mainstream algorithms. We first trained an outlier detection model to locate high-noise pixels in LDCT images. An image restoration model inspired by image inpainting was then developed to repair detected outliers, yielding LDCT images with spatially uniform noise. The preprocessed images were finally fed into conventional LDCT denoising algorithms for further processing. Evaluated on the public Mayo LDCT dataset, our preprocessing strategy increased the average PSNR by 1.5283 dB and SSIM by 0.8%, and reduced the RMSE by 4.911 HU when compared with raw LDCT images used as inputs for various denoising approaches. Thus, the Slide Region-based Convolutional Neural Network (SR-CNN) method provides a novel universal preprocessing approach for LDCT imaging by normalizing heterogeneous noise distribution, effectively enhancing image quality.