AI boosts identification of incident vertebral fractures on chest CT

A deep-learning algorithm that measures 3D whole thoracic vertebral bone mineral density on conventional chest CT can help predict incident vertebral fractures — offering an alternative to dual-energy x-ray absorptiometry, researchers have reported.

The study results could improve care of patients at risk of osteoporosis-related health conditions, wrote a team led by Quincy Hathaway, MD, PhD, of Johns Hopkins University in Baltimore, MD. Hathaway’s and colleagues’ findings were published March 11 in Radiology.

“Barriers to screening using dual-energy x-ray absorptiometry [DEXA] have greatly limited detection and treatment of osteoporosis, particularly in minority and underserved populations,” the authors explained. “Using opportunistic imaging, such as chest CT performed for other comorbid conditions, could help reduce fracture incidence as the prevalence of osteoporosis continues to increase.”

Overall prevalence of osteoporosis rose in the U.S. from 9.4% between 2007 and 2008 to 12.6% between 2017 and 2018, the authors noted. Bone mineral density (BMD) is typically measured using DEXA, and this modality can help predict risk of vertebral fractures. But its underuse and patient lack of adherence to screening remains a concern. That’s why using deep learning with chest CT — a common imaging exam — could offer an opportunity to assess BMD.

Hathaway’s group investigated whether TotalSegmentator, an nnU-net algorithm, could measure 3D vertebral body BMD across thoracic levels T1 to T10 on any noncontrast chest CT examination using data from a multicenter project called the Multi-Ethnic Study of Atherosclerosis (MESA), for which 2,956 participants underwent noncontrast chest CT with (n = 296) and without (n = 2,660) a phantom. For 594 patients, radiologist readers performed manual segmentation for T1 to T10 vertebrae on axial and sagittal planes, while TotalSegmentator produced 3D vertebral body segmentation of T1 to T10 levels with further postprocessing to remove cortical bone for all study participants. The researchers integrated vertebral BMD data with the Fracture Risk Assessment Tool (which had no BMD information) to predict incident vertebral fractures in 1,304 participants from a follow-up MESA exam, comparing the algorithm to the manual readings using Dice scoring (with 1 as reference).

The investigators reported the following:

  • Deep learning-derived 3D segmentations correlated with manual axial and sagittal segmentations (Dice score, 0.93 and 0.91 respectively).
  • Deep learning-derived 2D axial and sagittal BMD measurements had higher uncertainty compared with DL-derived 3D BMD measurements (average standard deviation, 2D axial and 2D sagittal versus 3D BMD: 65 mg/cm3 and 59 mg/cm3 compared to 41 mg/cm3; both p < 0.001).
  • Three-dimensional vertebral BMD with the Fracture Risk Assessment Tool showed better performance in predicting incident vertebral fractures compared with the tool’s use alone (area under the receiver operating characteristic curve [AUC] of 0.82 versus AUC of 0.66; p= 0.03).

Two-dimensional (2D) axial and sagittal segmentation overview. Representative noncontrast chest CT Digital Imaging and Communications in Medicine (A) axial and (B) sagittal segmentation images show manual labeling (white region of interest [ROI]) overlayed with inferred ROI by using the 2D deep learning–derived algorithm. (C) A custom instance segmentation ROI inference pipeline was developed using PixelLib (https://github.com/ayoolaolafenwa/PixelLib) and Mask R-CNN (https://github.com/matterport/Mask_RCNN). Representative noncontrast chest CT images show T10 examined by each step of the 2D deep learning–derived algorithm. In the inset box in C, the ROI alignment and deep learning process are illustrated, which results in a final segmentation (right side), where the inferred ROI is shown (purple area). Conv = convolution. Images and caption courtesy of the RSNA.Two-dimensional (2D) axial and sagittal segmentation overview. Representative noncontrast chest CT Digital Imaging and Communications in Medicine (A) axial and (B) sagittal segmentation images show manual labeling (white region of interest [ROI]) overlayed with inferred ROI by using the 2D deep learning–derived algorithm. (C) A custom instance segmentation ROI inference pipeline was developed using PixelLib (https://github.com/ayoolaolafenwa/PixelLib) and Mask R-CNN (https://github.com/matterport/Mask_RCNN). Representative noncontrast chest CT images show T10 examined by each step of the 2D deep learning–derived algorithm. In the inset box in C, the ROI alignment and deep learning process are illustrated, which results in a final segmentation (right side), where the inferred ROI is shown (purple area). Conv = convolution. Images and caption courtesy of the RSNA.

“This work is a crucial step toward the implementation of a fully automated platform in routine clinical practice to screen for bone health using any conventional chest CT examination; chest CT images are commonly obtained in individuals at risk of osteoporosis (eg, chronic obstructive pulmonary disease, coronary artery disease, congestive heart failure, and lung cancer screening),” the group concluded.

The study underscores the promise of opportunistic CT imaging, wrote Peter Steiger, PhD, now retired as CSO at Perceptive Imaging, in an accompanying editorial.

“Osteoporosis is insidious, and the first fracture is the most important to prevent,” Steiger noted. “We have the technology to identify those at risk and we have preventative treatments, which [is why] opportunistic screening for low BMD with chest and abdominal CT scans is so important.”

The complete study can be found here.