IEU-Net: Sequence learning of internal texture and external morphology for sonographic kidney segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nephrology

Background:

• Chronic kidney disease (CKD) prevalence is increasing globally.
• Renal ultrasound is crucial for kidney assessment.
• Automated kidney segmentation in ultrasound faces challenges due to limited data and quality variations.

Purpose of the Study:

• To develop an advanced deep learning model for automatic kidney segmentation in ultrasound images.
• To address data limitations and variability in multi-center datasets.
• To improve the accuracy and efficiency of kidney evaluation in clinical settings.

Main Methods:

• Developed IEU-Net, a sequence learning architecture integrating multi-center datasets.
• Utilized both original and modified nnU-Net components to learn internal texture and external morphology features.
• Trained and validated the model on diverse kidney ultrasound datasets from multiple countries.

Main Results:

• IEU-Net achieved a superior Dice Similarity Coefficient (DSC) of 0.8832.
• Demonstrated a 6.95% accuracy improvement over the standard nnU-Net.
• Showcased significant performance gains (3.64% to 20.13%) when integrated with various backbone networks.

Conclusions:

• IEU-Net shows significant potential for clinical application in kidney ultrasound segmentation.
• Deep learning-based segmentation can enhance diagnostic capabilities for CKD.
• The developed model and code are available for further research and development.