Investigating the Use of Traveltime and Reflection Tomography for Deep Learning-Based Sound-Speed Estimation in...

Area of Science:

• Medical Imaging
• Computational Imaging
• Biomedical Engineering

Background:

• Ultrasound computed tomography (USCT) is crucial for quantifying acoustic tissue properties like speed-of-sound (SOS).
• Full-waveform inversion (FWI) provides accurate SOS reconstruction but is computationally intensive.
• Deep learning-based image-to-image learned reconstruction (IILR) offers a computationally efficient alternative.

Purpose of the Study:

• To investigate the impact of input modalities on IILR for high-resolution SOS reconstruction in USCT.
• To compare the performance of dual-channel (traveltime tomography and reflection tomography) input against single-channel inputs.
• To evaluate the efficacy of IILR for SOS reconstruction using both numerical phantoms and clinical data.

Main Methods:

• A virtual USCT imaging system with numerical breast phantoms was utilized for systematic analysis.
• A supervised convolutional neural network (CNN) was trained to map dual-channel (TT and RT) inputs to high-resolution SOS maps.
• Single-input CNNs (TT or RT only) were trained for comparison, with accuracy assessed via NRMSE, SSIM, and PSNR.

Main Results:

• The dual-channel IILR method demonstrated superior performance compared to single-channel methods.
• Quantitative metrics (NRMSE, SSIM, PSNR) validated the improved accuracy and image quality of the dual-input approach.
• On clinical data, the dual-channel IILR achieved an ensemble average NRMSE of 0.2355, SSIM of 0.8845, and PSNR of 28.33 dB.

Conclusions:

• Combining traveltime tomography and reflection tomography as dual inputs significantly enhances high-resolution SOS reconstruction in USCT via IILR.
• This deep learning approach provides a computationally efficient and accurate alternative to FWI for USCT imaging.
• The findings suggest potential for improved diagnostic capabilities in breast imaging applications.