3:00 PM Abstract No. 61 Reducing cumulative radiation dose from computed tomography imaging in patients with drainage catheters: a retrospective study

Abstract

Inferior vena cava (IVC) filters are commonly encountered on radiographs of the abdomen, lumbar spine, and pelvis but are typically not identified in study reports according to their type and whether they are retrievable or not. We hypothesized that standard machine learning techniques including image analysis by convolutional neural networks (CNNs) would be able to differentiate between various types of IVC filters, and that these automatic classification methods could serve as a useful method to enrich registries of patients with IVC filters. Such patients may in turn be candidates for IVC filter retrieval. Materials: 530 radiographs of twelve categories of IVC filters were collected and regions of interest labelled manually with type of IVC filter. Standard image augmentation strategies were used to increase the number of training images. Transfer learning techniques were employed to retrain the Inception v3 convolutional neural network (CNN), part of the open source Tensorflow framework, to detect and classify the twelve categories of IVC filters. The accuracy of the retrained network was then evaluated by a set of 24 IVC filter radiographs not previously seen by the network. Results: The retrained network required 14 minutes to train on the augmented set of 5300 radiographs using CPU-only techniques and 8000 learning steps. Final test accuracy achieved was 80%, with training set accuracy up to 96%. In a set of 24 test images that the network had not previously seen, it correctly identified 75% of 12 classes of IVC filters. Confidence scores for correct identification of the filters were mainly clustered between 90-99% (range, 51-99%). The network was inaccurate on images of low brightness and contrast. The network could distinguish between similarappearing filters with high confidence, such as between OptEase and TrapEase filters, and between Tulip and Celect filters. Conclusions: In summary, we have trained a convolutional neural network to accurately identify and classify 12 classes of IVC filters. We anticipate that this will automate the detection and reporting of patients with IVC filters for purposes of registries and filter retrieval practices.