DeepUCT: Complex cascaded deep learning network for improved ultrasound tomography.

Area of Science:

• Medical Imaging
• Computational Imaging
• Biomedical Engineering

Background:

• Ultrasound computed tomography (UCT) offers radiation-free, cost-effective medical imaging.
• Image reconstruction in UCT, often using iterative methods like full-waveform inversion, is computationally intensive and ill-posed.
• Existing methods face challenges with computational burden and ill-posedness in nonlinear optimization.

Purpose of the Study:

• To develop a deep learning approach for near real-time ultrasound computed tomography image reconstruction.
• To overcome the computational expense and ill-posed nature of traditional iterative reconstruction methods.
• To directly learn the mapping from time-series sensor data to acoustic property images.

Main Methods:

• A deep learning model comprising two cascaded convolutional neural networks with an encoder-decoder architecture was developed.
• The model learns an intermediate mapping to extract knowledge for image reconstruction.
• The approach was validated using synthetic phantoms with simulated ultrasound data generated via k-wave toolbox.

Main Results:

• The deep learning method demonstrated robustness to noise in simulated data.
• Quantitative and qualitative results showed significant outperformance compared to state-of-the-art iterative methods.
• The proposed model achieved substantially reduced computational time compared to conventional full-waveform inversion.

Conclusions:

• Deep learning provides an effective solution for computationally demanding ultrasound computed tomography image reconstruction.
• The cascaded encoder-decoder network offers a robust and efficient alternative to traditional iterative algorithms.
• This advancement enables faster, more accurate clinical diagnosis through improved medical imaging techniques.