Learned Full Waveform Inversion Incorporating Task Information for Ultrasound Computed Tomography.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Ultrasound computed tomography (USCT) shows potential for breast imaging, but full-waveform inversion (FWI) reconstruction is computationally intensive.
• Accurate quantitative imaging of breast tissue acoustic properties is crucial for lesion detection.

Approach:

• A convolutional neural network (CNN) was developed to learn the mapping from USCT waveform data to speed of sound estimates.
• The CNN was trained using simulated USCT data from realistic numerical breast phantoms with a task-informed loss function.
• The model was validated against traditional FWI using metrics like RMSE, SSIM, and lesion detection performance.

Key Points:

• The CNN achieves accuracy comparable to FWI for image reconstruction (RMSE, SSIM).
• The AI model demonstrates superior performance in lesion detection tasks.
• Real-time image reconstruction is achieved, significantly reducing computational burden.

Conclusions:

• CNN-based reconstruction offers a computationally efficient alternative to FWI for USCT breast imaging.
• This AI-driven approach enhances diagnostic capabilities by improving lesion detection.
• The method holds promise for accelerating the clinical adoption of USCT for breast cancer screening.