Dilated Convolution ResNet with Boosting Attention Modules and Combined Loss Functions for LDCT Image Denoising

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Growing concerns about radiation exposure from computed tomography (CT) scans necessitate improved imaging techniques.
• Low-dose CT (LDCT) imaging offers reduced radiation but often suffers from noise, compromising diagnostic accuracy.
• Deep learning presents a promising avenue for denoising LDCT images while preserving essential diagnostic information.

Purpose of the Study:

• To propose a novel deep learning model for effective denoising of low-dose CT images.
• To investigate the impact of attention mechanisms on image denoising performance.
• To evaluate different loss functions for optimizing denoised image quality, focusing on detail preservation.

Main Methods:

• Development of a cascaded dilated residual network (ResNet) incorporating spatial and channel attention modules.
• Ablation studies comparing a basic ResNet with the proposed attention-augmented model.
• Evaluation of per-pixel, perceptual (VGG16-Net), and structural dissimilarity loss functions.
• Bench testing against state-of-the-art methods like BM3D, patch-GAN, and DRL-E-MP.

Main Results:

• Attention modules significantly improved the quality of denoised CT images compared to a standard ResNet.
• A combined loss function effectively prevented edge over-smoothing and enhanced textural and structural details.
• The proposed model demonstrated competitive or superior visual and quantitative results compared to existing state-of-the-art denoising techniques.

Conclusions:

• The proposed cascaded dilated ResNet with integrated attention modules offers a robust solution for LDCT image denoising.
• The strategic combination of loss functions is crucial for achieving high-fidelity denoised images that maintain diagnostic quality.
• This deep learning approach holds potential for reducing patient radiation exposure without sacrificing image quality in CT scans.