Background:
The aim of this study was to segment the kidney in ultrasound images at the sagittal and axial planes automatically. Then, predict kidney dimensions containing length, thickness, width, and volume of the kidney, parenchyma length, cortex length, and parenchyma volume.
Method:
We proposed a fast and accurate convolutional neural network, name Fast-Unet, to segment kidney and sinus in sagittal and axial planes of ultrasound images. Then, kidney length and thickness were measured from the resulting mask of sagittal frame and kidney width was measured from axial frame’s mask. Using these three dimensions, kidney volume was calculated. The exact procedure was performed for calculation of sinus volume. By subtracting renal volume and sinus volume the parenchyma volume was calculated.
Results:
The train-test dataset contained 1350 ultrasound images in sagittal and axial planes. The Dice and Jaccard coefficients were used to evaluate the segmentation step, and 98.3% and 94.8% for the sagittal frame and 93.2% and 90.8% for the axial frame were achieved respectively. The predicted values of renal measurement were also validated with radiologist’s report using root-mean-square-error (RMSE) metric and 0.09 cm, 0.08 cm, 0.08 cm, 0.04 cm, 0.01 cm, 1.3 cm3, 0.98 cm3, were achieved for kidney length, thickness, width, parenchyma length, the volume of kidney, and the volume of parenchyma respectively.
Conclusion:
To the best of our knowledge, this study was the first attempt to predict seven renal measurements using ultrasound images automatically. This can reduce the unnecessary referral to CT imaging and increase the precision of renal measurement using sonography.