Deep learning framework for automated frame selection in kidney ultrasound.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Ultrasound Technology

Background:

• Manual selection of kidney ultrasound frames is time-consuming and subjective.
• This variability can impact the reliability of clinical assessments.

Purpose of the Study:

• To develop and evaluate an automated deep learning framework for selecting optimal frames from kidney ultrasound videos.
• To enhance the efficiency and consistency of kidney ultrasound interpretation.

Main Methods:

• A dataset of 1,203 kidney ultrasound frames from 211 patients was curated and annotated.
• Several convolutional neural network models, including YOLOv11x-cls, were trained and compared for frame classification.
• The YOLO11x-cls model was optimized and evaluated using 5-fold patient-level cross-validation.

Main Results:

• The YOLO11x-cls model outperformed baseline architectures, achieving 100% F1-score for good quality frames.
• The framework attained an average cross-validation accuracy of 90% with minimal performance variance.
• The proposed method demonstrates robust and efficient automated best-frame selection.

Conclusions:

• The developed YOLO-based deep learning pipeline offers a promising solution for automated best-frame selection in kidney ultrasound.
• This technology can reduce manual effort and improve diagnostic reliability and reproducibility in clinical settings.