GEDFormer: Gradient Edge Detection in LDCT Image Denoising Transformer Model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Convolutional Neural Networks (CNNs) excel in feature extraction but struggle with global context in medical image denoising.
• Vision Transformers (ViTs) offer an alternative using self-attention to model both local and global image dependencies.
• Low-dose computed tomography (LDCT) image denoising is critical for reducing patient radiation exposure while maintaining diagnostic quality.

Purpose of the Study:

• To investigate a standalone Vision Transformer (ViT) framework for denoising low-dose computed tomography (LDCT) images.
• To introduce a self-guided gradient edge detecting attention module within the ViT framework.
• To evaluate the ViT-based denoising performance against established CNN and hybrid models.

Main Methods:

• Development of a ViT-based denoising framework incorporating a novel attention module.
• Rigorous evaluation using numerical data analysis and qualitative image inspection.
• Comparative analysis against state-of-the-art methods: BM3D, DSC-GAN, RED-CNN, and TED-Net.

Main Results:

• The ViT-based framework demonstrated superior performance in denoising LDCT images.
• The proposed attention module effectively preserved spatial and frequency details crucial for diagnostics.
• The standalone ViT approach showed competitive or improved results compared to CNN and hybrid models.

Conclusions:

• Standalone Vision Transformers provide a powerful framework for LDCT image denoising.
• The proposed method enhances diagnostic accuracy by preserving critical image information.
• ViTs represent a promising advancement in deep learning for medical image processing.