Self-supervised structural similarity-based convolutional neural network for cardiac diffusion tensor image...

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Machine Learning

Background:

• Cardiac Diffusion Tensor Imaging (DTI) enables non-invasive investigation of myocardial fiber structures.
• Low signal-to-noise ratio (SNR) in diffusion-weighted (DW) images limits accurate cardiac DTI.
• Effective noise reduction is crucial for reliable cardiac DTI analysis.

Purpose of the Study:

• To address limitations of existing denoising methods, such as redundancy dependence and over-smoothing.
• To propose a self-supervised learning model, SSECNN, for effective noise removal in cardiac DTI.
• To improve the accuracy and detail preservation in denoised DW images.

Main Methods:

• Developed a structural similarity-based matching algorithm to identify similar DW images.
• Utilized a convolutional neural network with residual blocks (SSECNN) for denoising.
• Implemented self-supervised training with similar noisy DW image pairs and an edge-weighted loss function to preserve details and avoid over-smoothing.

Main Results:

• SSECNN effectively reduced noise while preserving detailed texture and edge information in synthetic and real cardiac DTI datasets.
• Achieved significant improvements in RMSE, PSNR, and SSIM on synthetic data compared to state-of-the-art methods.
• Demonstrated enhanced SNR, CNR, and more regular helix and transverse angle maps on acquired cardiac DTI data.

Conclusions:

• The proposed SSECNN method effectively denoises cardiac DTI in a self-supervised manner by exploiting similarity between DW images.
• The edge-weighted loss function enhances detail preservation and mitigates over-smoothing.
• SSECNN overcomes limitations of existing methods, offering improved performance for cardiac DTI analysis.