Deep Learning-Assisted Quantitative Measurement of Thoracolumbar Fracture Features on Lateral Radiographs

Abstract

Objective: This study aimed to develop and validate a deep learning (DL) algorithm for the quantitative measurement of thoracolumbar (TL) fracture features, and to evaluate its efficacy across varying levels of clinical expertise.

Methods: Using the pretrained Mask Region-Based Convolutional Neural Networks model, originally developed for vertebral body segmentation and fracture detection, we fine-tuned the model and added a new module for measuring fracture metrics-compression rate (CR), Cobb angle (CA), Gardner angle (GA), and sagittal index (SI)-from lumbar spine lateral radiographs. These metrics were derived from six-point labeling by 3 radiologists, forming the ground truth (GT). Training utilized 1,000 nonfractured and 318 fractured radiographs, while validations employed 213 internal and 200 external fractured radiographs. The accuracy of the DL algorithm in quantifying fracture features was evaluated against GT using the intraclass correlation coefficient. Additionally, 4 readers with varying expertise levels, including trainees and an attending spine surgeon, performed measurements with and without DL assistance, and their results were compared to GT and the DL model.

Results: The DL algorithm demonstrated good to excellent agreement with GT for CR, CA, GA, and SI in both internal (0.860, 0.944, 0.932, and 0.779, respectively) and external (0.836, 0.940, 0.916, and 0.815, respectively) validations. DL-assisted measurements significantly improved most measurement values, particularly for trainees.

Conclusion: The DL algorithm was validated as an accurate tool for quantifying TL fracture features using radiographs. DL-assisted measurement is expected to expedite the diagnostic process and enhance reliability, particularly benefiting less experienced clinicians.

Keywords: Artificial intelligence; Deep learning; Radiography; Spinal curvatures; Spinal fractures; Spinal injuries.

Fig. 1.

Measurement of the compression rate and kyphotic angles based on 6-point labeling. Automated detection of a fractured vertebral body (VB) with a bounding box is demonstrated. Lines connecting the superior and inferior corner dots define the superior and inferior endplate lines, respectively. (A) Compression rate is calculated by the ratio of the reduced height of the fractured VB to the average height of the adjacent VBs (1–[height of fractured VB/mean height of adjacent VBs]). The blue numbers display the compression rates (%) of the anterior, middle, and posterior parts of the VB. The smallest value among these is selected as the representative compression rate. (B) Cobb angle is measured between the superior endplate of the upper adjacent VB and the inferior endplate of the lower adjacent VB. (C) Gardner angle is measured between the superior endplate of the upper adjacent VB and the inferior endplate of the fractured VB. (D) Sagittal index is calculated as the angle between the superior and inferior endplates of the fractured VB.

Fig. 2.

Sequential pipeline for deep learning-based quantitative analysis of TL fracture features. The diagram outlines the stepby- step workflow of the developed deep learning-based algorithm for analyzing quantitative features related to TL fractures. The pipeline includes data preprocessing, Mask R-CNN based vertebral body segmentation, and subsequent stages focused on measuring CR and kyphotic angles. RPN, Region Proposal Network; TL, thoracolumbar; Mask R-CNN, Mask Region-Based Convolutional Neural Networks; CR, compression rate.

Fig. 3.

Comparison of compression rate with 6-point labeling on thoracolumbar spinal radiographs by ground truth (GT) and a deep learning (DL) algorithm. Every visible vertebral body (VB) is marked with 6 points indicating the anterior, middle, and posterior columns. (A) Simple compression fracture. GT and DL labels nearly identical. (B) Failure of DL to place the middle upper point at the lowest site of the VB. (C) Incorrect placement of the middle pair by DL, suggesting difficulty in interpreting the 3-dimensional structure of a fractured VB from a nontrue lateral view 2-dimensional radiograph. The blue numbers indicate the compression rates (%) for the anterior, middle, and posterior parts of the VB.

Fig. 4.

Intraclass correlation coefficient plots between the deep learning algorithm and ground truth. ICC, intraclass correlation coefficient; CR, compression rate; CA, Cobb angle; GA, Gardner angle; SI, sagittal index; GT, ground truth; DL, deep learning.

Fig. 5.

Comparison of performance among the 4 readers for CR, CA, GA, and SI, without and with the DL assistance. An asterisk denotes a significant difference with p<0.05 by paired t-test. NS R2, neurosurgery second-year resident; Rad R2, radiology second-year resident; Rad R4, radiology fourth-year resident; DL, deep learning; ICC, intraclass correlation coefficient; CR, compression rate; CA, Cobb angle; GA, Gardner angle; SI, sagittal index; GT, ground truth.