A deep learning method for kidney segmentation in 2D ultrasound images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Accurate interpretation of 2D ultrasound (US) images is challenging.
• Portable 2D US devices with AI for organ identification are increasing, but effective methods are limited.
• Kidney segmentation in US images is crucial for diagnosis and monitoring.

Purpose of the Study:

• To evaluate the U-NET architecture for segmenting kidneys in 2D US images.
• To develop a strategy for creating a large, multi-view 2D US dataset from 3D US volumes.
• To assess the method's potential for clinical application in portable US systems.

Main Methods:

• Utilized the U-NET deep learning architecture for image segmentation.
• Generated a dataset of 3792 2D images by slicing 66 3D US volumes.
• Conducted experiments using the full dataset (WWKD) and a subset with kidney area > 500 mm2 (500KD).

Main Results:

• Achieved an average error of 2.88 ± 2.63 mm in the testing dataset.
• Demonstrated satisfactory performance in a proof-of-concept test on real 2D US images.
• The methodology showed potential for clinical relevance in portable US ecosystems.

Conclusions:

• The proposed deep learning method effectively segments kidneys in 2D US images.
• The strategy of using 3D US slices provides a valuable dataset for training AI models.
• This technique may enhance the interpretability of US images in clinical practice, particularly with portable devices.