Related Experiment Video
Updated: Jan 8, 2026
09:16
Supervised Machine Learning for Semi-Quantification of Extracellular DNA in Glomerulonephritis
Published on: June 18, 2020
7.3K
Deep Learning-Based Quantification of Residual Blood Clots in Single-Use Dialyzers Using Bedside Mobile-Captured
Yi-Ren Yeh1, Shih-Chun Yeh2, Ching-Fen Wu3
1Department of Mathematics, National Kaohsiung Normal University, Kaohsiung, Taiwan.
American Journal of Nephrology
|December 18, 2025
Summary
This study developed a machine learning model to detect blood clots in hemodialysis dialyzers using smartphone images. The AI model achieved higher accuracy than human experts, improving patient care.
Area of Science:
- Biomedical engineering
- Artificial intelligence in healthcare
- Medical imaging analysis
Background:
- Blood clot formation and dialyzer fiber blockage are critical issues in hemodialysis for end-stage kidney disease (ESKD) patients.
- Current methods for assessing dialyzer clots are often manual and subjective.
- Developing automated, objective methods for clot quantification is essential for improving patient outcomes and dialysis efficiency.
Purpose of the Study:
- To develop and validate a machine learning model for quantifying residual blood clots in dialyzer images.
- To utilize bedside smartphone cameras for image acquisition, enhancing accessibility and practicality.
- To improve the accuracy and robustness of blood clot detection in hemodialysis circuits.
Main Methods:
- Collected dialyzer images using mobile phones and applied preprocessing techniques including noise removal and image segmentation.
- Employed data augmentation to enhance model robustness and created composite images from both dialyzer ends.
- Developed a binary classification model using a pre-trained ConvNeXt architecture, incorporating Explainable AI (LIME) for clinical relevance assessment.
Main Results:
- The ConvNeXt model with pre-trained weights achieved an accuracy of 0.7572, outperforming the un-pretrained model (0.6971).
- The combined framework demonstrated the highest accuracy (0.7672) with improved robustness, indicated by a reduced standard deviation.
- The AI model's accuracy surpassed that of nephrology nurses (0.6271 and 0.6005) in manual clot level classification.
Conclusions:
- The developed ConvNeXt-based approach accurately detects residual blood clots in dialyzers using images from both ends.
- Explainable AI confirmed the model's focus on clinically relevant areas, ensuring reliable predictions.
- This efficient and robust AI model can aid clinical decision-making, reduce circuit downtime, and optimize anemia management in hemodialysis patients.
