Development and validation of deep learning models for identifying the brand of pedicle screws on plain spine radiographs

Yu-Cheng Yao; Cheng-Li Lin; Hung-Hsun Chen; Hsi-Hsien Lin; Wei Hsiung; Shih-Tien Wang; Ying-Chou Sun; Yu-Hsuan Tang; Po-Hsin Chou

doi:10.1002/jsp2.70001

Development and validation of deep learning models for identifying the brand of pedicle screws on plain spine radiographs

JOR Spine. 2024 Sep 17;7(3):e70001. doi: 10.1002/jsp2.70001. eCollection 2024 Sep.

Authors

Yu-Cheng Yao^{1

2}, Cheng-Li Lin³, Hung-Hsun Chen⁴, Hsi-Hsien Lin^{1

2}, Wei Hsiung^{1

2

5}, Shih-Tien Wang^{1

2

6}, Ying-Chou Sun^{1

7

8}, Yu-Hsuan Tang⁸, Po-Hsin Chou^{1

2}

Affiliations

¹ School of Medicine National Yang Ming Chiao Tung University Taipei Taiwan.
² Department of Orthopedics and Traumatology Taipei Veterans General Hospital Taipei Taiwan.
³ Department of Orthopaedic Surgery, National Cheng Kung University Hospital, College of Medicine National Cheung Kung University Tainan Taiwan.
⁴ Program of Artificial Intelligence and Information Security Fu Jen Catholic University New Taipei City Taiwan.
⁵ Department of Orthopedics Shin Kong Wu Ho-Su Memorial Hospital Taipei Taiwan.
⁶ Kinmen Hospital Ministry of Health and Welfare Kinmen Taiwan.
⁷ Department of Radiology Taipei Veterans General Hospital Taipei Taiwan.
⁸ Department of Medical Imaging and Radiological Technology Yuanpei University of Medical Technology Hsinchu Taiwan.

Abstract

Background: In spinal revision surgery, previous pedicle screws (PS) may need to be replaced with new implants. Failure to accurately identify the brand of PS-based instrumentation preoperatively may increase the risk of perioperative complications. This study aimed to develop and validate an optimal deep learning (DL) model to identify the brand of PS-based instrumentation on plain radiographs of spine (PRS) using anteroposterior (AP) and lateral images.

Methods: A total of 529 patients who received PS-based instrumentation from seven manufacturers were enrolled in this retrospective study. The postoperative PRS were gathered as ground truths. The training, validation, and testing datasets contained 338, 85, and 106 patients, respectively. YOLOv5 was used to crop out the screws' trajectory, and the EfficientNet-b0 model was used to develop single models (AP, Lateral, Merge, and Concatenated) based on the different PRS images. The ensemble models were different combinations of the single models. Primary outcomes were the models' performance in accuracy, sensitivity, precision, F1-score, kappa value, and area under the curve (AUC). Secondary outcomes were the relative performance of models versus human readers and external validation of the DL models.

Results: The Lateral model had the most stable performance among single models. The discriminative performance was improved by the ensemble method. The AP + Lateral ensemble model had the most stable performance, with an accuracy of 0.9434, F1 score of 0.9388, and AUC of 0.9834. The performance of the ensemble models was comparable to that of experienced orthopedic surgeons and superior to that of inexperienced orthopedic surgeons. External validation revealed that the Lat + Concat ensemble model had the best accuracy (0.9412).

Conclusion: The DL models demonstrated stable performance in identifying the brand of PS-based instrumentation based on AP and/or lateral images of PRS, which may assist orthopedic spine surgeons in preoperative revision planning in clinical practice.

Keywords: artificial intelligence; deep learning; instrumentation; pedicle screws.