Sensor-to-Image Based Neural Networks: A Reliable Reconstruction Method for Diffuse Optical Imaging of...

Area of Science:

• Biomedical Imaging
• Deep Learning
• Optical Tomography

Background:

• Deep learning is increasingly adopted for imaging tasks, including complex biomedical applications.
• Diffuse optical tomography (DOT) presents significant challenges due to its ill-posed and ill-conditioned nature.
• Existing regularization methods for DOT, like Tikhonov regularization (TR), have limitations.

Purpose of the Study:

• To develop a novel sensor-to-image neural network for diffuse optical imaging.
• To offer an alternative to traditional Tikhonov regularization (TR) methods.
• To create a deep learning model with enhanced capabilities for image reconstruction in DOT.

Main Methods:

• A sensor-to-image neural network combining multiple deep learning architectures was developed.
• The network incorporates an encoder for compressed input representation and a U-net with skip connections for feature extraction.
• A branching network structure was designed to accommodate ring-scanning measurement systems and diverse experimental data.

Main Results:

• The proposed network achieved reliable performance in both simulation and experimental cases.
• The model demonstrated superior performance compared to Tikhonov regularization (TR) and fully connected neural networks (FCNN).
• The network successfully reconstructed non-synthesized data and localized inclusions under various conditions.

Conclusions:

• The developed deep learning model is a feasible and promising alternative for diffuse optical imaging reconstruction.
• The strategy shows potential for clinical breast tumor imaging applications.
• The network's ability to handle diverse data and reconstruct complex scenarios highlights its clinical applicability.