Authors: Rashad Jabarkheel; Allison Park, BA; Chris Chute, BS; Pranav Rajpurkar, BS; Joe Lou; Katie Shpanskaya, BS; Lily Kim, BA; Emily McKenna; David Hong, MD; Thomas Wilson, MD; Kristen Yeom, MD (Mountain View, CA)
Introduction: CTA is the primary non-invasive diagnostic imaging modality for detecting cerebral aneurysms. However, aneurysm diagnosis by CTA can be challenging for clinicians due to lack of experience or subspecialty neuroradiology training, complex neurovascular anatomy, or the labor-intensive nature of identifying aneurysms. We aimed to develop and validate a novel 3D convolutional neural network architecture to automatically detect intracranial aneurysms on CTA and producelocation-specific segmentations. Methods: We retrospectively collected a dataset of 818 CTA head exams from 662 patients with 328 (40.1%) exams containing at least one clinically significant intracranial aneurysm (>3mm) and 490 (59.9%) exams (controls) without intracranial aneurysms between 2003 and 2017 at Stanford University Medical Center. We excluded exams with hemorrhage, ruptured aneurysm, post-traumatic or infectious pseudoaneurysm, arteriovenous malformation, and any surgical or endovascular hardware. The exams were split into a training set (611 exams, 494 patients) used to train our model, a development set (92 exams, 86 patients) used for model selection, andatest set (115 exams, 82 patients) to evaluate the final model’s performance. The ability of the model to augment clinician readers’ performancewas investigated with a crossover study design involving 8 clinical experts. Results: Clinicalexperts augmentedwith model segmentations had a statistically significant increase in both theirmicro-average sensitivity and accuracy. The clinical experts’ mean sensitivity increased from 0.831 to 0.890 and mean accuracy increased from 0.893 to 0.932. There was also an increase in inter-rater reliability among clinical experts, with an exact Fleiss’ kappa of 0.799 without augmentation and 0.858 with augmentation. The time to diagnosis for each clinical expert decreased on average by 5.96 seconds per case with augmentation. Conclusion: Our results suggest that segmentation models offer a promising approach for integrating deep learning into the clinical workflow and significantly improving human clinician diagnosis of intracranial aneurysms.