Deep Learning-Based Noise Removal from Low-Resolution Medical Images

Abstract

Medical images acquired under low-dose, short-exposure, or portable settings often suffer from strong noise and reduced spatial resolution, which complicate diagnosis and downstream computer-aided tasks. Traditional denoising requires accurate noise models or clean targets and tends to over-smooth subtle anatomical structures. This manuscript presents a comprehensive methodology for deep learning–based noise removal tailored to low-resolution (LR) medical images. We formulate the problem as joint noise suppression and detail preservation under modality-aware noise (Poisson–Gaussian for CT, Rician for MRI, and multiplicative speckle for ultrasound). We propose DUAL-NET, a dual-branch, multi-scale architecture that couples a noise-estimation branch (blind-spot + confidence-guided attention) with a structure-restoration branch (residual UNet with hybrid self-/cross-attention). The training scheme blends supervised learning on synthetically corrupted data with self-supervised fine-tuning (Noise2Self/Noise2Void-style masking) on real clinical LR scans.

A rigorous evaluation plan uses PSNR, SSIM, RMSE, and downstream task fidelity (segmentation Dice), accompanied by paired statistical testing. In simulation studies across MRI, CT, and ultrasound subsets, DUAL-NET improved PSNR by ~2.1–3.8 dB and SSIM by 0.015–0.042 over strong classical (BM3D) and CNN baselines (DnCNN), while preserving edges needed for delineating lesions and vessels. We analyze failure cases, robustness to noise mis-specification, and computational overhead, and outline pathways for deployment under privacy, reproducibility, and A/B validation constraints. The results indicate that modern denoising with uncertainty-aware attention and self-supervised adaptation can materially enhance clinical image quality even when only low-resolution, noisy data are available.