Low-Dose Computed Tomography (LDCT) is increasingly adopted for diagnostic imaging due to significant reduction in patient radiation exposure. However, this dose reduction inherently increases image noise, which substantially degrades image quality and compromises diagnostic accuracy. To address this trade-off, we propose a novel deep learning framework that integrates a Multi-Scale Residual Network (MSRN) with a Noise-Aware Attention mechanism. This architecture is designed to achieve superior noise suppression while critically preserving fine anatomical structures. The methodology employs self-supervised pretraining on unlabeled LDCT scans, to facilitate robust feature learning. We further incorporate adversarial training with a discriminator to ensure realistic image reconstruction, and utilize a hybrid loss functions to balance pixel-level accuracy, perceptual quality and visual reality. Experimental evaluations, quantified by Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) of LDCT images show significant performance gains in the quality of LDCT images with outputs as good as HDCT scans. This proposed approach offers a pathway to safer and more reliable clinical diagnosis by enabling correct image interpretation at substantially reduced radiation levels.
Parameswar et al. (Tue,) studied this question.