Yingying Lin1, Pek-Lan Khong1, Zhiying Zou2, Peng Cao3. 1. Diagnostic Radiology, The University of Hong Kong, Hong Kong, Hong Kong SAR, China. 2. Ultrasound Department, Shenzhen Maternity and Child Health Care Hospital, Shengzhen, China. 3. Diagnostic Radiology, The University of Hong Kong, Hong Kong, Hong Kong SAR, China. caopeng1@hku.hk.
Abstract
PURPOSE: Hydronephrosis is the dilation of the pelvicalyceal system due to the urine flow obstruction in one or both kidneys. Conventionally, renal pelvis anterior-posterior diameter (APD) was used for quantifying hydronephrosis in medical images (e.g., ultrasound, CT, and functional MRI). Our study aimed to automatically detect and quantify the fluid and kidney areas on ultrasonography, using a deep learning approach. METHODS: An attention-Unet was used to segment the kidney and the dilated pelvicalyceal system with fluid. The gold standard for diagnosing hydronephrosis was the APD > 1.0 cm. For semi-quantification, we proposed a fluid-to-kidney-area ratio measurement, i.e., [Formula: see text], as a deep learning-derived biomarker. Dice coefficient, confusion matrix, ROC curve, and Z-test were used to evaluate the model performance. Linear regression was applied to obtain the fluid-to-kidney-area ratio cutoff for detecting hydronephrosis. RESULTS: For regional kidney segmentation, the Dice coefficients were 0.92 and 0.83 for the kidney and dilated pelvicalyceal system, respectively. The sensitivity and specificity of detecting dilated pelvicalyceal system were 0.99 and 0.83, respectively. The linear equation was fluid-to-kidney-area ratio = (0.213 ± 0.004) × APD (in cm) for 95% confidence interval on the slope with R2 = 0.87. The fluid-to-kidney-area ratio cutoff for detecting hydronephrosis was 0.213. The sensitivity and specificity for detecting hydronephrosis were 0.90 and 0.80, respectively. CONCLUSION: Our study confirmed the feasibility of deep learning characterization of the kidney and fluid, showing an automatic pediatric hydronephrosis detection.
PURPOSE:Hydronephrosis is the dilation of the pelvicalyceal system due to the urine flow obstruction in one or both kidneys. Conventionally, renal pelvis anterior-posterior diameter (APD) was used for quantifying hydronephrosis in medical images (e.g., ultrasound, CT, and functional MRI). Our study aimed to automatically detect and quantify the fluid and kidney areas on ultrasonography, using a deep learning approach. METHODS: An attention-Unet was used to segment the kidney and the dilated pelvicalyceal system with fluid. The gold standard for diagnosing hydronephrosis was the APD > 1.0 cm. For semi-quantification, we proposed a fluid-to-kidney-area ratio measurement, i.e., [Formula: see text], as a deep learning-derived biomarker. Dice coefficient, confusion matrix, ROC curve, and Z-test were used to evaluate the model performance. Linear regression was applied to obtain the fluid-to-kidney-area ratio cutoff for detecting hydronephrosis. RESULTS: For regional kidney segmentation, the Dice coefficients were 0.92 and 0.83 for the kidney and dilated pelvicalyceal system, respectively. The sensitivity and specificity of detecting dilated pelvicalyceal system were 0.99 and 0.83, respectively. The linear equation was fluid-to-kidney-area ratio = (0.213 ± 0.004) × APD (in cm) for 95% confidence interval on the slope with R2 = 0.87. The fluid-to-kidney-area ratio cutoff for detecting hydronephrosis was 0.213. The sensitivity and specificity for detecting hydronephrosis were 0.90 and 0.80, respectively. CONCLUSION: Our study confirmed the feasibility of deep learning characterization of the kidney and fluid, showing an automatic pediatric hydronephrosis detection.