Unsupervised Deep Learning Framework for Low-Dose CT Denoising

A study published on arXiv presents an unsupervised deep learning approach aimed at denoising low-dose computed tomography (CT) images. While low-dose CT minimizes radiation exposure, it also introduces noise that can complicate diagnoses. The proposed framework, drawing from Cycle-GAN, utilizes a U-Net for extracting features at multiple scales, incorporates an attention mechanism for fusing features, and employs a residual network for transforming these features. To tailor the network to the specifics of medical imaging, perceptual loss is utilized. The researchers created a genuine low-dose CT dataset and established an extensive training regimen. This technique seeks to enhance image quality without the need for paired noisy-clean training datasets, making it suitable for clinical applications.