APU-Net: an attention-based physical U-Net model improves the image quality of DOT reconstruction by reducing DOT image artifacts and improving target depth profile.
by Minghao Xue (https://opticalultrasoundimaging.wustl.edu/)
Traditional diffuse optical tomography (DOT) reconstructions suffer from image artifacts due to various factors such as the proximity of DOT sources to shallow lesions, poor optode-tissue coupling, tissue heterogeneity, and large high-contrast lesions causing shadowing effects. This study introduces an attention-based U-Net (APU-Net) model with Contextual Transformer (CoT) attention modules to enhance DOT image quality and improve lesion diagnostic accuracy. Trained on simulation and phantom data, and evaluated on clinical data, the APU-Net model effectively reduced artifacts by 26.83% on average and significantly improved image contrast in deeper regions, with increases of 20.28% and 45.31% for the second and third target layers, respectively. These improvements demonstrate the potential of the APU-Net model in enhancing DOT reconstructions for better breast cancer diagnosis.
Dataset and model can be found here
- Python: 3.7+
- torch(pytorch): 1.10+
- torchvision: 0.11.1+
- numpy: 1.21.2+
- scipy: 1.7.1+
- scikit-learn: 1.3+
Please email Minghao Xue at [email protected] if you have any concerns or questions.