Abstract No. 481 Training a convolutional neural network to detect refractory variceal bleeding in cirrhotic patients

Purpose: To evaluate the diagnostic accuracy of three convolutional neural networks (CNN) for classifying CT findings of variceal bleed requiring portosystemic intervention. Materials: From January 2009 to January 2018, a HIPAA‐compliant, institutional review board–approved, retrospective clinical study used analyze contrast‐enhanced abdominopelvic CT scans from 73 patients who experienced variceal hemorrhage requiring TIPS/RTO and 91 patients without variceal hemorrhage but with pressure gradient proven portal hypertension. 3 CNN were developed with a combination of different training data sources [custom CNN, transfer learning model (VGG16, ImageNet), and transfer learning model (VGG16, ImageNet) with augmentation]. 82 cases were randomly assigned for total training, 20 cases were allocated for validation, and an additional 20 for testing. Models were assessed by testing validation accuracy and area under the receiver operating characteristic curve (AUC). Results: Preliminary data demonstrated a non‐random pattern of accuracy and detectability. The custom CNN model reported a validation accuracy of 0.65 and AUC of 0.79. The pretrained CNN model (VGG16, ImageNet) demonstrated an accuracy of 0.55 and AUC of 0.63. The pretrained CNN with augmentation demonstrated an accuracy of 0.65 and AUC of 0.71. Conclusions: Preliminary data from CNN models can identify CT findings of variceal bleed requiring portosystemic intervention with a non‐random accuracy (AUC, 65‐70%). Performance and generalization of CNNs can be enhanced by using transfer learning with datasets enriched with labeled medical images.