Deep-Learning-Driven High Spatial Resolution Attenuation Imaging for Ultrasound Tomography (AI-UT)

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Quantitative ultrasound (QUS) and ultrasound tomography (USCT) estimate breast tissue attenuation but have limitations in spatial resolution and noise susceptibility.
• Current methods like QUS reduce spatial resolution for attenuation mapping, while USCT's full wave inversion is prone to background noise.
• Accurate high-resolution attenuation imaging is crucial for detailed breast tissue characterization and potential cancer detection.

Purpose of the Study:

• To develop a deep learning (DL) based method for high-resolution, low-variance ultrasonic attenuation imaging of the breast.
• To enhance DL model performance by integrating spatial correlation between speed of sound (SOS) and attenuation.
• To validate the DL-generated attenuation images for improved breast segmentation and potential use as a diagnostic biomarker.

Main Methods:

• Utilized RF data from 60 angle views of the QTI Breast Acoustic CT (BACT) scanner as input for a DL model.
• Implemented a DL approach that generates attenuation images as output.
• Incorporated spatial correlation between SOS and attenuation as a constraint to improve DL model performance and image quality.

Main Results:

• The DL-based method successfully generated high-resolution attenuation images with improved quality when SOS structural information was included.
• The inclusion of SOS data enhanced the DL model's performance.
• DL-generated attenuation images demonstrated validated structural information and attenuation values comparable to literature and SOS-based segmentation.

Conclusions:

• The proposed DL method significantly improves spatial resolution and reduces variance in breast ultrasonic attenuation imaging.
• Integrating SOS structural information enhances the accuracy and performance of the DL-based attenuation estimation.
• High-resolution attenuation images generated by DL can serve as a valuable additional biomarker for breast cancer diagnosis and enable better breast segmentation.