Detecting and Localizing Cervical Spine Fracture Using an End-To-End 3D Deep Learning Pipeline

Purpose: Over 1.5 million spine fractures occur annually in the United States alone, resulting in over 17,000 spinal cord injuries. The most common site of spine fracture is the cervical spine, and computer tomography (CT) is the modality of choice for detecting these fractures. Commercial computer-aided diagnosis tools are being used to triage studies and prioritize patients with a suspected cervical spine fracture (CSFx). However, they come short when the patient has superimposed degenerative disease and osteoporosis and are reported to have variable sensitivity, ranging from 54% to 76%. This study aims to develop a deep learning algorithm that can robustly detect and localize cervical spine fractures on CT scans.

Methods/Materials: Data from the RSNA-ASNR Cervical Spine Fracture Challenge (n=2019) was used to train and validate the algorithm. This data was gathered from twelve different institutions that can lead to a more generalizable model. Inspired by the holistic view of the radiologists, an end-to-end 3D approach was used to prevent mistaking artifacts that mimic fracture in 2D slices (e.g., nutrition canals). We first used a weakly supervised approach to localize C1-C7 vertebrae by applying a trained model on the data and then trained a segmentation model to finetune the segmentation masks. Finally, each 3D vertebrae volume was resized to 256x256x64 and used to train a 3D EfficientNetB4 to binary classify images as with or without fracture.

Results: The segmentation model reached a Dice Similarity Coefficient of 0.92. The vertebrae-level fracture detection model had an area under the receiver operating curve (AUROC) of 0.92. When ensemble, on a patient level, the model reached an AUROC of 0.93 with a sensitivity of 87% and specificity of 77%.

Conclusions: This work presents a 3D deep learning model to detect and localize CSFx in CT scans from multiple institutions. The high sensitivity of the model makes it a suitable candidate for screening. Additionally, the per-vertebrae nature of the workflow enables precise localization of the fracture, which can help speed up the evaluation by expert radiologists to verify or reject the model’s prediction.