AutoKV-Net: Calculating Single Kidney Volume in Two-Dimensional Ultrasound Automatically by Mimicking Sonographers
Measuring single kidney volume in ultrasound may be a useful surrogate biomarker for renal function, as the renogram requires scarce materials and high costs. Measuring volumes using ultrasound is a promising approach, but requires expertise to identify and measure. In this pilot study, we explored a machine learning approach to SKV estimation in 2D ultrasound cine clips in transplant and native kidneys. A machine learning algorithm was trained using 514 images that were annotated by the sonographers for the kidney capsule. A different set of 16 patients (32 cine clips, 16 volumes) were randomly selected. The sonographers measured the kidney lengths and widths in these cines. The trained algorithm processed the cines, and generated a prediction for the frame with maximum capsule area. It fit a bounding box to it, where the recentangle dimensions serve as the kidney dimensions. The annotators SKV measured was 220 ± 119 mL (95% CI: 162, 278), while the algorithm SKV measured was 224 ± 109 mL (95% CI: 171, 277) with no significant differences. The mean absolute difference was 4 mL while the mean relative difference was 6.7%. For comparison, a +0.1 cm error in each of the algorithm’s measured dimensions using the same volume equation would result in an absolute difference of 78 mL and a relative difference of 42%. There was no performance difference between native kidneys and transplanted kidneys. Manual measurements require 5 to 10 minutes, whereas the algorithm requires < 10 seconds. In this pilot study, an automatic algorithm to measure SKV from 2D ultrasound cines was demonstrated to be comparable to expert estimates of volume. Such a technique may permit for the automatic, rapid and reliable measurement of SKV using conventional standard 2D ultrasound imaging. Future work includes correlation against renogram functional tests and prospective validation of the technique.