PENet - A Scalable Deep-Learning Model for Automated Diagnosis of Pulmonary Embolism Using Volumetric CT Imaging

Abstract

Background: Pulmonary embolism (PE) is a life-threatening clinical problem and computed tomography pulmonary angiography (CTPA) is the gold standard for diagnosis. Prompt recognition of the diagnosis and immediate treatment is critical for controlling morbidity and mortality rates while PE remains among the diagnoses most frequently missed or delayed, in part due to lack of radiologist availability, diagnostic error, and delayed imaging results communication to clinical providers. Methods: We developed a deep learning model - PENet, to detect PE on volumetric CTPA scans without any manual intervention. The PENet is a 77-layer 3D convolutional neural network that has been pre-trained on the Kinetics-600 dataset and fine-tuned only on a retrospective CTPA dataset collected from a single academic institution. The PENet model was evaluated in detecting PE on the hold-out test set (198 CTPA studies) from the same institution as well as from the external test set (227 CTPA studies) collected from an external institution. Findings: PENet achieved an AUROC of 0.83 on automatically detecting central PE on the hold out test set and 0.84 on external institution's dataset. The results represent successful application of 3D convolutional neural network models for the complex task of PE diagnosis and demonstrates sustained performance on data from an external institution. Interpretation: Our model could be applied as a triage tool in ED to automatically identify clinically important central PEs allowing for prioritization for diagnostic radiology interpretation and improved care pathways via more efficient diagnosis. Funding Statement: Research reported in this publication was supported by the National Library of Medicine of the National Institutes of Health under Award Number R01LM012966, Stanford Child Health Research Institute (Stanford NIHNCATS-CTSA Grant #UL1 TR001085), and Philips Healthcare. Declaration of Interests: The authors state: No conflict of interest, financial or other, exists. Ethics Approval Statement: All the patient data used in this study are de-identified and the study was approved by the Institutional Review Board (IRB).