AutoSOUL: Deep Convolutional Neural Network for Autonomous Selection of Ultrasound Speckle Tracking Parameters

Figures

• 

  Fig. 1.. Demonstration of the nonlinear increments within the α1 and α2 ranges (reported in Table I) per dataset, plotted on a log scale, as a function of the ordinal number when sorted from minimum to maximum α1 and α2 values per dataset.

• 

  Fig. 2.. flowchart summarizing the strain image labeling process. Th1 and Th2 denote the minimum and maximum thresholds for the acceptable SNR values, whereas Th3 and Th4 refer to the minimum and maximum thresholds for the acceptable SR values.

• 

  Fig. 3.. CNN architecture employed in AutoSOUL.

• 

  Fig. 4.. Summary of the AutoSOUL framework.

• 

  Fig. 5.. (a) Example of strain images from the single-inclusion simulation test case, obtained with varying α1 and α2 parameter combinations and corresponding CNN prediction probabilities. (b) CNN prediction probability distribution from the single-inclusion simulation test case as a function of log-α1 and log-α2 values.

• 

  Fig. 6.. Axial strain results obtained with the single-inclusion simulated dataset. The white and black boxes indicate target and background ROIs, respectively, utilized to calculate one of the 30 SNR measurements (black ROI only), one of the 120 CNR, and one of the 120 SR measurements. The arrows indicate the directions of ROI displacement to obtain the remaining SNR, CNR, and SR values.

• 

  Fig. 7.. Axial strain results obtained with the multi-inclusion simulated dataset. The white and black boxes indicate two target ROIs and one background ROI, respectively, utilized to calculate one of the 30 SNR measurements (black ROI only), two of the 120 CNR, and two of the 120 SR measurements. The arrows indicate the directions of ROI displacement to obtain the remaining SNR and CNR values.

• 

  Fig. 8.. Axial strain results obtained with the public phantom dataset. The B-mode image is shown with 40-dB dynamic range. The blue and red boxes indicate target and background ROIs, respectively, utilized to calculate one of the 30 SNR measurements (blue ROI only), one of the 120 CNR, and one of the 120 SR measurements. The arrows indicate the directions of ROI displacement to obtain the remaining SNR, CNR, and SR values.

• 

  Fig. 9.. Axial strain results obtained with the JHU phantom dataset. The B-mode image is shown with 40-dB dynamic range. The blue and red boxes indicate target and background ROIs, respectively, utilized to calculate one of the 30 SNR measurements (blue ROI only), two of the 120 CNR, and two of the 120 SR measurements. The arrows indicate the directions of ROI displacement to obtain the remaining SNR, CNR, and SR values.

• 

  Fig. 10.. Axial strain results obtained with the in vivo breast mass dataset. The B-mode image is shown with 50-dB dynamic range. Yellow arrows indicate the mass boundary [13]. The blue and red boxes indicate target and background ROIs, respectively, utilized to calculate one of the 30 SNR measurements (blue ROI only), one of the 120 CNR, and one of the 120 SR measurements. The arrows indicate the directions of ROI displacement to obtain the remaining SNR, CNR, and SR values.