An in-silico simulation study to generate computed tomography images from ultrasound data by using deep learning...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Ultrasound has limited sensitivity for parenchymal lesion detection compared to contrast-enhanced computed tomography (CT).
• Developing advanced imaging techniques is crucial for improving diagnostic accuracy and patient outcomes.

Purpose of the Study:

• To investigate the potential of using raw ultrasound data to generate CT-like images via deep learning.
• To enhance lesion detection capabilities by improving image quality and reducing artifacts.

Main Methods:

• Utilized k-wave and Astra simulation toolkits to generate ultrasound and CT datasets.
• Trained a pix2pix conditional Generative Adversarial Network (cGAN) on simulated phantom data.
• Evaluated image quality using generalized contrast-to-noise ratio (gCNR) and Structural Similarity Index (SSIM).
• Assessed radiologist performance and inter-observer agreement using the Jaccard Index for lesion segmentation.

Main Results:

• CT-like images generated from ultrasound data showed significantly improved gCNR and SSIM.
• Lesions undetectable in B-mode ultrasound were sometimes highlighted in the generated CT-like images.
• Radiologist performance in lesion segmentation improved significantly with the addition of machine learning-augmented B-mode images.
• Inter-observer agreement among radiologists also showed significant improvement.

Conclusions:

• Deep learning models can effectively translate ultrasound data into CT-like images.
• This translation improves image quality, enhances lesion detectability (e.g., by alleviating shadowing), and increases inter-observer agreement.
• The cGAN architecture enables successful generation of improved diagnostic images even with relatively small datasets.