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) is a promising breast imaging technique.
• Full-waveform inversion (FWI) provides high-resolution quantitative images but is computationally intensive.
• High computational cost limits the clinical application of FWI for breast imaging.

Purpose of the Study:

• To investigate the use of a convolutional neural network (CNN) for real-time USCT image reconstruction.
• To develop an AI-driven method that reduces the computational burden of FWI.
• To assess the accuracy and lesion detection performance of the CNN compared to FWI.

Main Methods:

• A CNN was trained using supervised learning with a task-informed loss function.
• Training utilized a large dataset of simulated USCT measurements from realistic numerical breast phantoms (NBPs).
• Performance was evaluated against FWI using RMSE, SSIM, and lesion detection metrics on a hold-out dataset.

Main Results:

• The CNN achieved accuracy comparable to FWI in terms of RMSE and SSIM.
• The CNN demonstrated superior performance in lesion detection tasks.
• Real-time image reconstruction was achieved, significantly reducing computational time.

Conclusions:

• Supervised learning with CNNs offers a viable alternative to traditional FWI for USCT breast imaging.
• The proposed CNN method accelerates image reconstruction without compromising diagnostic accuracy.
• This AI approach has the potential to enhance the clinical utility of USCT for breast cancer detection.