Abstract: TH-OR26
Using a Smartphone Camera at the Bedside to Detect and Quantify Residual Blood Clots in Single-Use Dialyzers with Paired-Image Classification
Session Information
- Augmented Intelligence and Digital Health Advances
October 24, 2024 | Location: Room 2, Convention Center
Abstract Time: 05:40 PM - 05:50 PM
Category: Augmented Intelligence, Digital Health, and Data Science
- 300 Augmented Intelligence, Digital Health, and Data Science
Authors
- Lin, Hugo Y.-H., Kaohsiung Medical University Chung Ho Memorial Hospital, Kaohsiung, Taiwan
- Yeh, Yi-Ren, National Kaohsiung Normal University, Kaohsiung, Taiwan
Background
Blood clot formation and subsequent blockage of the capillary fibers of dialyzers is challenging for patients with end-stage kidney disease (ESKD) on hemodialysis. This study aimed to develop a machine learning detection model that utilizes the information obtained from dialyzer images to detect blood clots of dialyzers early.
Methods
We gather dialyzer images captured at the bedside using mobile devices daily. Initially, we segment the image to extract the dialyzer part and filter out background noise unrelated to the residual blood clot analysis. Subsequently, we apply data augmentation to enhance robustness. Capturing snapshots at both ends of the dialyzer provides two distinct perspectives merged into a single composite image. Our classification task involves two classes: instances with less than 10% residual blood clots (666 samples) and those with 30% residual blood clots (538 samples). Due to the limited image dataset, we utilize a pre-trained image classification model, EfficientNet, and fine-tune it using our collected data.
Results
In our experiment, we strategically split the dataset into training (60%), validation (20%), and testing (20%) sets. Conducting ten random trials provided insights into the model's stability and consistency. The proposed model achieved an average recognition rate of 76.33%, surpassing human evaluators who scored 62.71% and 60.05% on the testing set. It demonstrates the model's superior performance and potential to enhance recognition capabilities in relevant domains. The results prove the viability of employing image-based approaches for detecting blood clots in dialyzers.
Conclusion
Our proposed framework, including data preprocessing and model training, exhibits potential for applications in various clinical scenarios where recognition relies on information derived from images. This approach involves using image data to amalgamate information from both ends of the dialyzer. By employing the EfficientNet model to detect blood clots, this integrated approach significantly improves blood clot detection in dialyzers compared to human evaluators.
Funding
- Government Support – Non-U.S.