FMUnet: Frequency Feature Enhancement Multi-level U-Net for Low-Dose CT Denoising with a Real Collected LDCT Image Dataset

Accompanying the widespread use of CT systems in medical diagnostics has highlighted concerns about the health risks associated with X-ray radiation exposure. Despite reducing the use of X-rays, low-dose computed tomography (LDCT) as a method to mitigate radiation risk is often plagued by quantum noise due to the scarcity of X-ray photons in low-dose scenarios. This results in image edge discontinuities, smoothing of small target structures, and the emergence of low-contrast visual effects. These manifestations of visual degradation primarily occur within the high-frequency band of the image, this study focuses on enhancing the quality of LDCT images by optimizing the utilization of frequency domain features. Specifically, we adopt a multi-level supervised U-shaped neural network and introduce a novel Frequency Feature Attention (FFA) mechanism. FFA utilizes convolution to diversify frequency features, then modulates them using channel weights to enhance learning of beneficial frequencies. We also introduce frequency domain loss based on fast Fourier transform to supervise the model's learning in the frequency domain. Furthermore, considering that synthetic data might introduce biases or distribution mismatches absent in real data, we established a real LDCT dataset. For each volunteer, one regular-dose CT scan and one low-dose CT scan are conducted respectively, resulting in a total of 4310 pairs of NDCT-LDCT images. Through experiments on the contributed dataset, our method produces superior results and outperforms other methods, significantly improving the quality of low-dose CT images, and providing strong technical support for reducing X-ray radiation risks while ensuring the accuracy of image diagnosis.