Improving diffuse optical tomography reconstruction using an attention-based U-Net post-processing framework

Area of Science:

• Biomedical imaging
• Optical physics
• Medical image processing

Background:

• Diffuse optical tomography (DOT) is a noninvasive imaging technique with significant biomedical potential.
• Conventional DOT reconstruction algorithms suffer from ill-posedness, resulting in poor spatial resolution, quantitative accuracy, and robustness.
• There is a need for advanced methods to improve DOT image quality.

Purpose of the Study:

• To propose an attention-enhanced deep learning post-processing method, ART-U-Net-CBAM, to improve DOT image reconstruction.
• To evaluate the performance of ART-U-Net-CBAM against conventional methods using simulations and phantom experiments.
• To demonstrate the effectiveness of attention mechanisms in enhancing DOT image quality.

Main Methods:

• Developed ART-U-Net-CBAM, combining the algebraic reconstruction technique (ART) with a U-Net incorporating a convolutional block attention module (CBAM).
• Trained the network exclusively on simulated DOT data.
• Validated the method using numerical simulations and phantom experiments with circular and elliptical targets.

Main Results:

• ART-U-Net-CBAM significantly outperformed ART and ART-U-Net in reconstruction accuracy and noise robustness.
• The proposed method demonstrated superior spatial resolution and structural similarity compared to baseline methods.
• Quantitative evaluations confirmed the enhanced performance across various targets.

Conclusions:

• Attention-enhanced deep learning post-processing is an effective strategy for improving DOT image quality.
• ART-U-Net-CBAM offers a generalizable approach to overcome the inherent limitations of conventional DOT reconstruction.
• The findings support the clinical translation of advanced DOT imaging techniques.