A KL Divergence-Based Loss for In Vivo Ultrafast Ultrasound Image Enhancement with Deep Learning.

Area of Science:

• Medical imaging
• Ultrasound technology
• Artificial intelligence in medicine

Background:

• Ultrafast ultrasound imaging achieves high frame rates but suffers from low image quality.
• Convolutional neural networks (CNNs) show promise for enhancing ultrasound image quality without sacrificing speed.
• Existing CNN models often perform poorly on in vivo images due to training on simulated data.

Purpose of the Study:

• To develop and validate a CNN-based method for enhancing single plane wave (PW) ultrasound acquisitions using in vivo training data.
• To introduce a novel training loss function addressing the dynamic range and echogenicity distributions of radio frequency data.
• To rigorously evaluate the performance of the proposed method on a large in vivo dataset.

Main Methods:

• Development of a CNN model trained on a dataset of 20,000 in vivo ultrasound images.
• Implementation of a specialized training loss function incorporating Kullback-Leibler divergence for echogenicity distribution preservation.
• Comparative analysis using peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) against coherent compounding of 87 PWs.

Main Results:

• The CNN method significantly improved image quality, increasing PSNR from 16.47 ± 0.80 dB to 20.29 ± 0.31 dB.
• SSIM improved from 0.105 ± 0.060 to 0.272 ± 0.040, indicating better structural preservation.
• The method effectively reduced artifacts in ultrafast ultrasound images.

Conclusions:

• The proposed CNN approach, trained on in vivo data with a tailored loss function, substantially enhances ultrafast ultrasound image quality.
• This method offers a viable solution for improving diagnostic accuracy in real-time ultrasound applications.
• The findings highlight the potential of AI in overcoming limitations of current ultrasound imaging techniques.