Deep learning assisted sparse array ultrasound imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Ultrasound Technology

Background:

• Sparse array ultrasonography faces challenges with grating lobe artifacts and reduced image resolution.
• Deep learning offers potential solutions for reconstructing high-quality ultrasound images from limited data.

Purpose of the Study:

• To develop and validate a deep learning predictive model for restoring grating lobe artifacts and enhancing image resolution in sparse array ultrasonography.
• To assess the performance of a deep learning-assisted sparse array imaging system using reduced channel counts (64 and 16 out of 128).

Main Methods:

• A deep learning model was trained using sparse array (64- and 16-channel) ultrasound images as input and dense array (128-channel) images as ground truth.
• The system was demonstrated on ex vivo porcine teeth, evaluating image quality metrics and gingival thickness measurements.
• Generalization capability of the deep learning model was also investigated.

Main Results:

• The deep learning model significantly improved image quality metrics (SSIM, MSE, PSNR) compared to standard sparse array images (p < 0.0001).
• Reconstructed images achieved resolutions close to the ground truth (0.18 mm and 0.15 mm vs. 0.15 mm).
• Gingival thickness measurements showed high agreement between predicted and ground truth images (bias -0.01 to 0.02 mm, Pearson's r = 0.99) and clinical probing (<0.05 mm bias).

Conclusions:

• A deep learning-assisted sparse array system can reconstruct high-resolution ultrasound images using significantly fewer channels (down to 16/128).
• The model demonstrated generalization capabilities for the 64-channel array, with further optimization needed for the 16-channel array.
• This approach holds promise for improving the efficiency and quality of ultrasound imaging in dental and potentially other applications.