Deep Learning-Based Denoising With Noise Power Spectrum and CycleGAN in Low-Dose Head CT Images

Head Computed Tomography (CT) is a crucial diagnostic tool in neurological imaging. While low-dose computed tomography (LDCT) aims to minimize radiation exposure, it presents challenges in balancing noise reduction with detail preservation, particularly for head CT where small anatomical structures are vital. This study proposes a deep learning-based denoising technique to enhance the quality of head LDCT images and address the limitations of traditional denoising approaches that often compromise diagnostic accuracy. Our method utilizes noise power spectrum (NPS) and CycleGAN to generate synthetic noisy data for unsupervised model training. Results demonstrate that our approach effectively improves LDCT image quality. Quantitative analysis shows that the root mean square error (RMSE) was below 6.92 Hounsfield Units (HU). The visual quality also indicates prominent improvement in image quality. This novel technique offers a promising solution for enhancing LDCT images in neurological diagnostics, potentially enabling wider adoption of low-dose protocols without compromising diagnostic accuracy.