Generative Adversarial Network With Robust Discriminator Through Multi-Task Learning for Low-Dose CT Denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Reducing radiation dose in computed tomography (CT) is crucial for minimizing secondary cancer risk.
• Low-dose CT (LDCT) images often suffer from increased noise, potentially hindering diagnostic accuracy.
• Existing deep learning methods for LDCT denoising face challenges in visual congruence, performance metrics, and robustness.

Purpose of the Study:

• To develop an advanced deep learning model for effective LDCT image denoising.
• To address limitations of current methods, including visual discrepancies and performance variability.
• To enhance the robustness and diagnostic utility of LDCT images.

Main Methods:

• A novel generative adversarial network (GAN) incorporating multi-task learning in its discriminator.
• Introduction of restoration consistency (RC) and non-difference suppression (NDS) for improved discriminator capabilities.
• Integration of residual fast Fourier transforms with convolution (Res-FFT-Conv) blocks for enhanced feature representation.

Main Results:

• The proposed GAN demonstrated superior denoising performance across quantitative and qualitative evaluations.
• Multi-task learning in the discriminator provided more effective feedback for generator denoising.
• The model achieved better results compared to state-of-the-art denoising techniques in visual scoring by radiologists.

Conclusions:

• The novel GAN architecture with multi-task learning and specialized blocks significantly improves LDCT denoising.
• The proposed regulatory mechanisms enhance GAN training and discriminator performance.
• This approach offers a promising solution for high-quality LDCT imaging, aiding in accurate diagnoses and reduced radiation exposure.