Back to Fall Congress

Background: Ultrasound is the mainstay of imaging for pediatric hydronephrosis, yet is limited by its subjective assessment, absence of a consistently interpreted grading system, and lack of correlation with functional imaging modalities. In order to improve the clinical utility of ultrasound imaging for hydronephrosis, we developed a semi-automatic quantitative image analysis algorithm. The primary goal was to identify thresholds of safety for hydronephrotic renal units where diuretic renography could be avoided based on washout times. Our underlying hypothesis was that renal morphology is correlated with renal drainage and can be quantitated using computer-based image analysis, to define clinically relevant thresholds without the use of diuretic renography in a significant fraction of patients.
Methods: After IRB approval a test cohort of 51 hydronephrotic patients with concurrent renal 2D ultrasound imaging and diuretic renography was selected. Manual segmentation of renal parenchyma and collecting system in both normal and hydronephrotic renal units was performed to create a control population for program calibration and algorithm development. A separate population of normal and hydronephrotic renal units, also with ultrasound imaging and diuretic renography was examined using machine learning and shape models of the kidney and collecting system were developed. A total of 131 morphological parameters were computed, including renal parenchymal and collecting system size, and shape descriptors including parameters of curvature. These were then applied to the test cohort using machine learning techniques, such as support vector machines, to identify ultrasound-based safety thresholds that agree with the clinical parameters of the diuretic renogram (half-time for washout).
We then set threshold levels of washout times that would be clinically relevant from 20 to 40 minutes. A best-fit model was then derived for each threshold using optimal morphological parameters to categorize the renal units. Receiver operating characteristic (ROC) curve analysis was performed and sensitivity, specificity and area under the curve (AUC) were determined.
Results: Optimal results were obtained by means of a non-linear support vector machine classification system. The design parameters of the model were adjusted in order to maximize the sensitivity of detecting severe cases of hydronephrosis; that is, no case with a washout time above the threshold was misclassified. For each one of the five washout thresholds considered, (20 to 40 minutes), and at 100% sensitivity (all severe cases of hydronephrosis correctly identified), the specificities of the method (percentages of patients in the safe zone) varied from 62 to 85%. The corresponding AUC values were between 0.85 and 0.91.
Conclusions: Quantitative image analysis of renal ultrasounds in children with hydronephrosis can identify thresholds of clinically significant washout times with very high sensitivity and clinically acceptable specificity in order to reduce the number of diuretic renograms in up to 60% of children who would otherwise routinely undergo these studies. This technology can be further developed to permit more robust assessment of hydronephrosis to facilitate clinical decision-making and communication, and reduction in invasive imaging with ionizing radiation.