Dual-Branch GAN-Driven Super-Resolution for Low-Dose CT Image Enhancement under Radiological Noise Constraints
Judy Simon, Nellore Kapileswar
Abstract
Reducing radiation in patients with LDCT is common, but the procedure often leads to poor image quality. A new system, called Dual-Branch Generative Adversarial Network (GAN), is presented in this paper to improve low-dose computed tomography (LDCT) images in the face of complex radiological noise. The network has two paths, one for restoring texture and one for recovering the most important edges of organs. Each branch is merged by an attention-guided feature aggregation module that adjusts the role of each stream depending on its surroundings. An NMU that is aware of the nature of CT noise is included to mimic and reduce Poisson-Gaussian noise. This makes the model stronger and able to generalize well when used with several CT scanners. Besides, a type of loss function is created by combining perceptual, adversarial and a new radiologist consistency loss, using expert-provided structural knowledge to improve the enhancement process. Experiments on both NIH DeepLesion and Mayo LDCT datasets suggest that the proposed model is more effective than current methods in terms of both technical scores and the quality of results viewed by radiologists. The framework proposes a way to generate reliable, high-resolution CT images while using much less radiation, making diagnostic procedures both safer and more effective in clinical settings.