Investigating the Use of Convolutional Neural Networks for Prenatal Hydronephrosis Ultrasound Image Classification

Abstract

Prenatal hydronephrosis is a common condition that involves the accumulation of urine with consequent dilatation of the collecting system in fetal infants. There are several hydronephrosis classifications, however, all grading systems suffer from reliability issues as they contain subjective criteria. The severity of hydronephrosis impacts treatment and follow up times and can therefore directly influence a patient’s well-being and quality of care. Considering the importance of accurate diagnosis, it is concerning that no accurate, reliable or objective grading system exists. We believe that developing a convolutional neural network (CNN) based diagnostic aid for hydronephrosis will improve physicians’ objectivity, inter-rater reliability and accuracy. Developing CNN based diagnostic aid for ultrasound images has not been done before. Therefore, the current thesis conducted two studies using a database of 4670 renal ultrasound images to investigate two important methodological considerations: ultrasound image preprocessing and model architecture. We first investigated whether image segmentation and textural extraction are beneficial and improve performance when they are applied to CNN input images. Our results showed that neither preprocessing technique improved performance, and therefore might not be required when using CNN for ultrasound image classification. Our search for an optimal architecture resulted in a model with 49% 5-way classification accuracy. Further investigation revealed that images in our database had been mislabelled, and thus impacted model training and testing. Although our current best model is not ready for use as diagnostic aid, it can be used to verify the accuracy of our labels. Overall, these studies have provided insight into developing a diagnostic aid for hydronephrosis. Once our images and their respective labels have been verified, we can further optimize our model architecture by conducting an exhaustive search. We hypothesize that these two changes will significantly improve model performance and bring our diagnostic aid closer to clinical application.