Abstract No. 596 Determining suitability of inferior vena cava (IVC) filter removal with intravascular ultrasound (IVUS): retrospective review of experience in 11 patients

Abstract

performed using PyTorch and 2 NVIDIA K80 GPUs. Receiver operating characteristic (ROC) curves with area under the curve (AUC) and standard diagnostic measures (e.g., sensitivity, specificity, accuracy) were used to evaluate the DCNN's performance. Results: The DCNN achieved AUC of 0.96 for detection of IVC filter. At optimal diagnostic thresholds, the DCNN achieved accuracy of 90.9% and sensitivity and specificity of 88.2% and 93.7%, respectively. Conclusions: A DCNN trained on a small set of images from similar but different imaging modalities (radiographs and fluoroscopic images) are able to automatically detect IVC filters with reasonable diagnostic performance. Our study demonstrates proof-of-concept of deep learning towards this task, which will likely be improved with the use of larger datasets as well as extending towards other types of Interventional devices and hardware.