Self-adaption and texture generation: A hybrid loss function for low-dose CT denoising.

Area of Science:

• Medical Imaging
• Deep Learning
• Image Denoising

Background:

• Deep learning models for low-dose CT (LDCT) denoising often rely on per-pixel loss functions (e.g., MAE, MSE).
• These traditional methods overlook regional variations in denoising difficulty, leading to loss of crucial texture information in CT images.

Purpose of the Study:

• To develop a hybrid loss function that adaptively addresses varying noise levels across different regions of CT images.
• To improve the diagnostic value of LDCT images by balancing denoising and texture preservation.

Main Methods:

• A novel hybrid loss function combining weighted patch loss (WPLoss) and high-frequency information loss (HFLoss) was proposed.
• WPLoss adaptively adjusts loss weights based on regional denoising difficulty, improving upon MAE for challenging areas.
• HFLoss specifically targets and preserves image texture details by analyzing high-frequency information.

Main Results:

• The proposed hybrid loss function demonstrated improved denoising performance across multiple deep learning models.
• Quantitative metrics such as peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) were significantly enhanced.
• Visual assessment confirmed effective noise suppression and superior retention of image details compared to existing methods.

Conclusions:

• The developed hybrid loss function offers improved denoising for LDCT images, enhancing the performance of existing deep learning models.
• Validation across diverse datasets and models confirmed its strong generalization capabilities.
• This approach enables the acquisition of high-quality, low-radiation CT images, supporting accurate disease diagnosis while minimizing patient exposure.