DCT-UNet: a UNet architecture for diffuse correlation tomography.

Area of Science:

• Biomedical Optics
• Medical Imaging
• Machine Learning

Background:

• Diffuse correlation tomography (DCT) is an optical imaging technique for assessing tissue blood flow.
• Conventional blood flow index (BFI) reconstruction methods face challenges due to ill-posed mathematical problems, impacting accuracy and robustness.
• Existing algorithms rely on light electric field temporal autocorrelation functions, which are sensitive to noise and data sparsity.

Purpose of the Study:

• To develop a deep learning framework for enhanced DCT image reconstruction.
• To establish a robust mapping between optical signals and blood flow tomographic images.
• To overcome the limitations of conventional DCT reconstruction methods, enabling faster and more accurate blood flow imaging.

Main Methods:

• A novel UNet architecture, termed DCT-UNet, was proposed, incorporating deformable convolutions and gated units.
• The DCT-UNet utilizes a group aggregation bridge (GAB) for improved encoder-decoder connections.
• Deep supervision from UNet++ was integrated for multi-scale mask generation, enhancing the loss function and GAB input.

Main Results:

• The DCT-UNet network demonstrated superior accuracy and robustness in computer simulations and phantom experiments.
• The deep learning approach effectively addressed ill-posed problems inherent in DCT reconstruction.
• The proposed method allows for rapid blood flow imaging, outperforming traditional iterative techniques.

Conclusions:

• The DCT-UNet network represents a significant advancement in DCT-based blood flow imaging.
• This deep learning framework overcomes key limitations of conventional methods, offering improved performance.
• DCT-UNet shows strong potential for future applications in both physiological research and clinical settings.