Hybrid Neural Networks for Precise Hydronephrosis Classification Using Deep Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Urology

Background:

• Hydronephrosis assessment using renal ultrasound images can be subjective and variable.
• Accurate segmentation of kidney and fluid is crucial for reliable hydronephrosis diagnosis.

Purpose of the Study:

• To develop and evaluate a deep learning framework for automatic kidney and fluid segmentation in renal ultrasound images.
• To enhance diagnostic accuracy and reduce variability in hydronephrosis assessment.

Main Methods:

• A dataset of 1731 renal ultrasound images was used for training and evaluation.
• A deep learning framework with DenseNet201, Feature Pyramid Network (FPN), and Self-Organizing Neural Network (SelfONN) layers was implemented.
• Segmentation performance was evaluated using Dice coefficient, precision, and recall; hydronephrosis classification used the fluid-to-kidney area ratio.

Main Results:

• The framework achieved high segmentation performance for kidneys (Dice: 0.92) and fluid (Dice: 0.89).
• Hydronephrosis classification accuracy reached 94% using the fluid-to-kidney area ratio.
• The model demonstrated robust performance across varied image qualities.

Conclusions:

• The developed framework provides an automated and objective pipeline for renal ultrasound image analysis.
• The system facilitates standardized and reproducible hydronephrosis assessment with high accuracy.
• Future work includes model optimization and integration of explainable AI for clinical use.