A Novel Self Attention Deep Learning-based Image Denoising Model for Low Dose CT Images

Computed tomography (CT) is a powerful medical imaging technique that provides cross-sectional images of the body. It uses X-rays and computer processing to create detailed pictures of the internal structures of the body. One of the key advantages of CT is its ability to produce images of soft tissues, bones, and blood vessels with high resolution and clarity. CT scans are commonly used in emergency medicine, oncology, neurology, and many other medical specialties. The need for clear imaging often exposes patients to significant doses of harmful radiation, especially in cases of multiple CT scans or elevated radiation dosages. Due to radiation concerns, low-dose CT (LDCT) has become a major area of research. LDCT exposes patients to only 25% of the total amount of radiation used in conventional full-dose CT scans. Although LDCT has radiation reduction benefits, it produces images with higher noise and artifacts. These inherent noise and artifacts make LDCT images less suitable for clinical use, reducing their diagnostic utility. In contrast, full-dose CT scans use standard radiation levels to generate high-quality images of the internal structures of the body. Current denoising methods focus on quantitative metrics rather than realistic image preservation, leading to over-denoised or blurred CT images. This research project introduces a new pipeline that combines deep learning methodologies and traditional image processing filters (such as wavelet transform and empirical mode decomposition) to effectively denoise LDCT images and generate CT images with reduced noise. Our aim is to balance diagnostic accuracy with patient safety.