Deep Learning Ultrasound Computed Tomography Under Sparse Sampling.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Artificial Intelligence

Background:

• Ultrasound computed tomography (USCT) shows promise for breast tumor detection and characterization.
• Quantitative speed of sound (SOS) and acoustic attenuation imaging are key USCT applications.
• High-quality USCT reconstruction often requires expensive, high-transducer-count systems or suffers from low quality with sparse arrays.

Purpose of the Study:

• To develop a high-quality quantitative SOS reconstruction framework for USCT under sparse sampling conditions.
• To improve image quality and computational efficiency compared to existing methods.
• To enable cost-effective and faster breast tumor detection using USCT.

Main Methods:

• A novel framework combining the bent-ray algorithm with a convolutional neural network (SRSS-Net) was proposed.
• SRSS-Net was designed for efficient image quality improvement from sparsely sampled USCT data.
• The method was evaluated using synthetic and real breast tissue datasets, including an inhomogeneous phantom.

Main Results:

• SRSS-Net demonstrated superior performance in artifact suppression, structural preservation, and quantitative accuracy compared to state-of-the-art methods.
• The proposed network achieved significantly faster computation speeds.
• Fine-tuning strategies were identified as beneficial for real-world applications.

Conclusions:

• SRSS-Net offers a promising solution for high-quality, quantitative SOS reconstruction in sparse-sampling USCT.
• The method has significant potential for real-time breast cancer detection due to its imaging and computational performance.
• Further validation and fine-tuning are recommended for clinical implementation.