Deep Learning of Renal Scans and Kidney Ultrasounds in Children with Antenatal Hydronephrosis - Project Summary/Abstract The candidate for this K08 Career Development Award aims to become a leader and innovator in applying bioinformatics and computational research to the study of pediatric urologic diseases with the goal of improving patient outcomes. Antenatal hydronephrosis (ANH) is one of the most common anomalies identified on prenatal ultrasound, found in up to 4.5% of all pregnancies. Up to 30% of ANH is secondary to urinary tract obstruction, most often unilateral ureteropelvic junction obstruction (UPJO). Early identification of infants with obstruction is critical, as obstructive uropathy exposes the kidney parenchyma to high pressures which causes progressive loss of kidney function. Patients with ANH are surveilled with kidney ultrasounds; a mercaptoacetyltriglycerine (MAG3) diuretic renal scan is performed to evaluate for UPJO when there is high suspicion on ultrasound. However, current evidence has established a wide range of “indeterminate” drainage times (approximately 10- 40 minutes), for patients with ANH and suspected UPJO, making interpretation and application of renal scan findings difficult. The resulting high inter-observer variability among pediatric urologists, can result in unnecessary surgery in infants without obstruction or delay treatment for those with obstruction. The proposed project aims to utilize machine learning to derive clinically useful anatomic signatures of the kidney from MAG3 renal scans and kidney ultrasound imaging data and diagnose UPJO with greater accuracy than the current standard of care. In an initial cohort of patients, his preliminary model built using only features from MAG3 renal scans identified clinically significant obstruction in patients with suspected UPJO. In Aim 1 of the proposed study, he will develop a multimodality deep learning model combining features from both MAG3 renal scans and kidney ultrasounds to predict clinically significant obstruction in children with suspected UPJO. He will compare the performance of the multimodality model to his prior model built using only MAG3 renal scans. In Aim 2, he will compare the best performing model from Aim 1 to that of an expert panel of pediatric urologists. Finally, in Aim 3 he will provide his best performing model’s output to physicians during their clinical decision-making process and determine if the model improves physician confidence and accuracy when predicting clinically significant obstruction in patients with suspected UPJO. The result of this work will be a model that can augment a pediatric urologist’s surgical decision making for patients with UPJO. The proposed project will provide the candidate the opportunity to develop skills in implementation science methodology, advanced machine learning methods, advanced statistics, trial design and execution, and development as a leader. The candidate’s primary mentor, Chris Flask, PhD, is an international expert in quantitative imaging analysis techniques and has an outstanding track-record of mentorship. The training the candidate receives during the award coupled with his advanced research degree, exceptional mentorship, robust institutional research infrastructure, and experience in machine learning research will position him to achieve scientific independence and write an R01 during the K08 award.
