E2109. Deep Learning System for Automated Detection of Posterior Ligamentous Complex Injury in Patients with Thoracolumbar Fracture on MRI
Authors
Sang Won Jo;
Hallym University Dongtan Sacred Heart Hospital
Eun Kyung Khil;
Hallym University Dongtan Sacred Heart Hospital
Seun Ah Lee;
Hallym University Dongtan Sacred Heart Hospital
Jihe Lim;
Hallym University Dongtan Sacred Heart Hospital
Yu Sung Yoon;
Soonchunhyang University Bucheon Hospital
Jae Hyeok Lee;
DEEPNOID
Objective:
To develop a deep learning algorithm for automated detection and localization of posterior ligamentous complex (PLC) injury in patients with thoracolumbar (TL) fracture on MRI and evaluate its diagnostic performance.
Materials and Methods:
In a retrospective and multicenter study, a midline sagittal fat-suppressed T2-weighted MR image with PLC injury based on radiologic reports were extracted. The deep learning algorithm development in this study was conducted through two major steps. The first step was to train the deep learning algorithm (attention U-net) to segment the entire soft tissue area (background anatomy) including two PLCs (normal or damaged PLC) above and below the spinous process of TL fracture. The second step was to have the deep learning algorithm detect and classify the PLC injury area expressed as a T2 high signal intensity lesion in the segmented image by the attention U-net. To carry out the first and second steps, spine MR images were randomly divided into two data sets: training set of 300 examinations with TL fracture (150 PLC injury group and 150 PLC normal group) and internal test set of 100 examinations with TL fracture (50 PLC injury group and 50 PLC normal group). External test data set also consisted of 100 examinations (50 PLC injury group and 50 PLC normal group) with TL fracture, which were extracted based on radiological reports from a different institution. After manual ground truth segmentation of TL fracture, background anatomy, PLC injury, a deep learning algorithm based on inception-ResNet V2 architecture was established with the training set. Its dice score, sensitivity, specificity and AUC were evaluated in the internal and external test sets.
Results:
The dice score of attention U-net in the first step was 0.849 for the internal test set and 0.773 for the external test set. The sensitivity, specificity, and AUC of inception-ResNet V2 architecture in the second step was 88%, 82% and 0.928, for the internal test set and 86%, 74% and 0.916 for the external test set, respectively.
Conclusion:
A deep learning algorithm detected PLC injury in patients with TL fracture with a high diagnostic performance which was validated using external data set.
