Deep-Learning for High Quality and High Quantitative Ultrasonic Echo Imaging.

Area of Science:

• Medical imaging
• Artificial intelligence
• Ultrasound technology

Background:

• Traditional ultrasonic imaging faces challenges with artifacts like multiple echoes and lobe echoes.
• Accurate segmentation and characterization of breast tumors are crucial for differential diagnosis.
• Deep learning (DL) offers potential solutions for image enhancement and automated analysis in medical imaging.

Purpose of the Study:

• To investigate the application of deep learning (DL) for improving ultrasonic (US) echo imaging quality.
• To evaluate DL models for artifact reduction, image enhancement, and superresolution in US imaging.
• To explore the use of DL for segmenting and differentiating benign and malignant breast tumors.

Main Methods:

• Simulations were conducted using deep learning algorithms.
• DL models were trained for tasks including multiple echo reduction, lobe echo grading, and wave separation.
• Image processing techniques such as US attenuation correction and superresolution were implemented.
• Segmentation of breast tumors (benign and malignant) was performed using DL.

Main Results:

• DL successfully reduced multiple echoes and graded lobe echoes in US images.
• Separation of multiply crossed waves and US attenuation correction imaging were achieved.
• Superresolution imaging of reflection and scattering was demonstrated.
• Accurate segmentation of benign and malignant breast tumors was performed.

Conclusions:

• Deep learning significantly enhances the quality and quantitative accuracy of ultrasonic echo imaging.
• DL-based segmentation holds potential for automatic differential diagnosis of in vivo human breast tumors.
• This study highlights the clinical relevance of DL in advancing breast cancer diagnostics.