Award: Stewart B. Dunsker, MD Award
Authors: Michael Zhang, MD; Lily Kim; Robin Cheong; Ben Cohen-Wang; Katie Shpanskaya; Jessica Wetstone; Nidhi Manoj; Pranav Rajpurkar; Kristen Yeom (Stanford, CA)
Cervical spine trauma accounts for over 1 million emergency department visits each year in the Northern America. Accurate and timely diagnosis is crucial as delay in intervention may lead to further injury and even paralysis or death. However, detection of cervical spine fractures on computed tomography (CT) scans may be challenging, especially for subtle fractures and adult patients with preexistent degenerative changes. Automated detection model can help clinicians make faster and better decisions in time-critical settings.
A total of 1347 CT scans obtained at our institution were divided into a training set (990 normal, 222 with fractures) and a validation set (98 normal, 37 with fractures). Manual annotation of all scans containing confirmed cervical fractures were performed by a board-certified neuroradiologist. For the model-building, ResNet-101, a 3D convolutional neural network, was used to extract fracture-related features from each image. With the Feature Pyramid Network architecture, predictions from each image were combined to make a prediction for the entire scan. Focal Loss function penalized high-confidence false predictions, helping the model focus on pertinent features and disregard easier negatives such as the background.
Model performance was measured with Area Under the Receiver Operating Characteristic (AUROC) and Area Under Precision-Recall Curve (AUPRC) metrics, with the expert human annotation serving as the ground truth. Performance was assessed at both the image- (AUROC 0.87, AUPRC 0.52) and scan-level (AUROC 0.85, AUPRC 0.82). Examples of common errors included over-prediction of C2 fractures and under-prediction of fractures in less common locations.
We present the first deep-learning model trained end-to-end to automatically detect fractures in the cervical spine, validated with the largest collection of cervical spine CT scans to our knowledge. The promising performance of our model in predicting cervical fractures is suggestive of its utility in neurosurgical practices for improved diagnostic quality.