Page 89 - Read Online

Glaser et al. Art Int Surg. 2025;5:1-15 https://dx.doi.org/10.20517/ais.2024.36 Page 7

(reference variability analysis, etc.)
standard) against manual
measurements
Kim et al. 1,807 lateral X-ray Mask R-CNN 0.4°-3.0° NA Mean absolute error Compared to Did not include 200 test images; Mask R-CNN 18 36 epochs
[24]
2023 radiographs; variety of for vertebral vs. manual measurements by images with statistical analysis (ResNet 101 images
spinal conditions segmentation measurements: 3 surgeons severe spinal (MAE, ICC, etc.) backbone) per
0.4°-3.0° for (criterion deformities or against manual batch
parameters; dice standard) implants measurements
similarity coefficient:
92.6% for
segmentation

Yeh et al. 2,210 whole spine X-ray Cascaded Landmark: NA Median error: 1.75- Compared to Single center 400 test images; Cascaded NA 120 epochs
[25]
2021 lateral radiographs; pyramid 1.75-3.39 3.39 mm for measurements by data; did not statistical analysis pyramid net (early
variety of spinal network + mm; landmarks; 3 doctors (ground include images against ground stopping
conditions differentiable parameter: parameter errors: truth) with vertebral truth applied)
spatial to 0.1°-6.6° mean 0.1°-6.6°, anomalies measurements
numerical median 0.03-5.3°
transform layer
Orosz et 600 lateral spine X-ray CNN for Not 0.75-0.92 Intraclass correlation Compared to Single-center Statistical analysis Convolutional NA NA
[26]
al. 2022 radiographs for segmentation + reported coefficient between measurements by data for (ICC, mean error, NN + U-Net
training; 200 lumbar U-Net for AI and human raters: expert human validation; did etc.) against
spine radiographs (100 landmark 0.85-0.92 pre-op, raters not assess intra- manual
pre-op, 100 post-op) detection 0.75-0.91 post-op rater reliability measurements by
for testing expert raters
Gami et al. 100 images to train X-ray YOLO version 3 Cobb: NA Average absolute Compared to Testing only on Cadaver testing + YOLOv3 CNN NA NA
[27]
2022 model, 130 images to CNN 1.726° difference - Cobb radiographic single cadaver verification testing
test model angle: 1.726°, plumb measurements in model and on artificial
line: 0.415 cm cadaver model artificial templates
templates
Schwartz 816 lateral lumbar X-ray MultiResUNet ≤ 4.6° NA Mean absolute Compared to 10% failure rate 163 test images; MultiResUNet NA NA
et al. radiographs including CNN + difference vs. measurements by for Cobb angle; statistical analysis
[28]
2021 some with computer vision surgeons: ≤ 4.6° for 3 orthopedic spine potential for against manual
instrumentation/hip pipeline parameters; success surgeons measurement surgeon
prostheses rate: 90%-100% skew measurements
Aubert et 68 biplanar X-ray CNN for Landmark: NA Mean error: 1.6-2.3 Compared to Small dataset Comparison to CNNs NA NA
[29]
al. 2019 radiographs with anatomical 1.6-2.3 mm; mm for landmarks; expert supervised from single multiple expert
variety of spinal landmark parameter: 2.8°-4.7° for spinal reconstructions center supervised
conditions detection to fit 2.8°-4.7° parameters; 1°-2.1° (ground truth) reconstructions;
statistical spine for pelvic automated vs.
model parameters expert agreement
analysis

Nguyen et 500 whole spine lateral X-ray Decentralized 1.156°- ≥ 0.8 for 10 of Correlation Compared to Difficulty with 30 test images + VGG-net Batch 50 epochs
[30]
al. 2022 radiographs with CNN 6.318° 12 parameters coefficient: ≥ 0.8 for manual parameters statistical analysis based CNN size: 32
variety of conditions 10 of 12 parameters; measurements by related to T1 against standard architecture
mean absolute error: experienced vertebrae; reference