Abstract: TH-OR142
Automatic Classification of Arteriovenous Fistula Aneurysms Using Artificial Intelligence
Session Information
- Vascular Access and Complications of Hemodialysis
November 07, 2019 | Location: 201, Walter E. Washington Convention Center
Abstract Time: 05:18 PM - 05:30 PM
Category: Dialysis
- 704 Dialysis: Vascular Access
Authors
- Zhang, Hanjie, Renal Research Institute, New York, New York, United States
- Sor, Murat, Azura Vascular Care, Woodland Park, New Jersey, United States
- Koh, Elsie, Azura Vascular Care, Woodland Park, New Jersey, United States
- Preciado, Priscila, Renal Research Institute, New York, New York, United States
- Maddux, Dugan, Fresenius Medical Care North America, Waltham, Massachusetts, United States
- Usvyat, Len A., Fresenius Medical Care North America, Waltham, Massachusetts, United States
- Larkin, John W., Fresenius Medical Care, Waltham, Massachusetts, United States
- Kotanko, Peter, Renal Research Institute, New York, New York, United States
Background
Trained professionals examine arteriovenous fistulas and diagnose aneurysm (AVFA) stages, but there is no fully automated classification method. Our goal is to build a convolutional neural network (CNN) to automatically classify AVFA stages.
Methods
We collected 15-20 sec “panning” videos from 30 patients with two AVFA categories, 23 patients with AVFA stage 2 (enlarged AVF with hypopigmented skin) and 7 patients with AVFA stage 3 (enlarged fistula with open ulcer). The videos were collected against a white background. Patients had diverse skin tones. We extracted the video frames that then comprised our image set. Each image has three color channels, the image shape was 960 x 540 (Figure 1). We used 80% of the patients’ videos for CNN training and the remainder for validation using the Amazon SageMaker machine learning platform (Amazon, Seattle, Washington).
Results
We trained the CNN in two modes. First, the transfer learning mode, utilized the pre-trained Amazon SageMaker image classification algorithm, fine-tuned to our image data set. Second, the full training mode, where the CNN was trained from scratch with our image data set. With both of these training modes, we were achieved a > 90% classification accuracy in the validation images.
Conclusion
CNN is able to automatically classify AVFA. Automation of that process is expected to reduce workload, provide timely AVFA diagnosis, and improve patient care.
Figure 1: A: AVFA with enlarged fistula and hypopigmented skin, B: AVFA with enlarged fistula and ulceration.
Funding
- Commercial Support –