Using Deep Learning Algorithms to Grade Hydronephrosis Severity: Toward a Clinical Adjunct.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pediatric Nephrology

Background:

• Hydronephrosis severity grading from renal ultrasound images is subjective.
• Deep learning offers an objective, data-driven approach to image classification.
• Convolutional neural networks (CNNs) are a type of deep learning algorithm suitable for image analysis.

Purpose of the Study:

• To evaluate the potential of a CNN for grading hydronephrosis severity using ultrasound images.
• To classify images according to the 5-point Society for Fetal Urology (SFU) system.
• To explore the application of CNNs in developing clinical decision and teaching aids.

Main Methods:

• A five-layer CNN was developed to analyze 2,420 sagittal hydronephrosis ultrasound images from 673 pediatric patients.
• The CNN was trained for five-way classification (all grades) and two-way classification (low vs. high grades; grades II vs. III).
• Performance was evaluated using accuracy and weighted F1 scores.

Main Results:

• In five-way classification, the CNN correctly classified images within one grade of the provided SFU label 94% of the time.
• The model achieved 51% accuracy with a weighted F1 score of 0.49 for exact five-way classification.
• For low vs. high grade classification, accuracy was 78% (F1=0.78), and for grades II vs. III, accuracy was 71% (F1=0.71).

Conclusions:

• A CNN approach is applicable and effective for classifying hydronephrosis ultrasound images.
• The developed CNN model performs significantly above chance and can classify images with high concordance to expert grading.
• Further research into deep learning-based clinical adjuncts for hydronephrosis assessment is warranted.