Page 87 - Read Online

Page 6 Glaser et al. Art Int Surg. 2025;5:1-15 https://dx.doi.org/10.20517/ais.2024.36

Table 1. Main table describing study characteristics

Mean Comparison Neural
Paper Dataset details Imaging Model details absolute Correlation Accuracy metrics with other Key Validation network Batch No. of
size
epochs
limitations
approach
coefficient
error methods architecture
Chae et al. Training - 400; X-ray Decentralized 1.45°-3.52° NA Mean absolute error: Compared to Requires 40 test Custom NA Initial:
[18]
2020 resolution - 3,240 × CNN; multiple 1.45°-3.52° for manual multiple ordered radiographs; decentralized 0.001,
1,080 pixels; variety - orders parameters measurement by datasets, comparison to CNN SGDM
57% normal spine, experienced training time; manual momentum
20% lumbar lordosis, surgeons, as well limited diversity measurements by 0.95
24% thoracic kyphosis as regression experienced
CNN model surgeons
Wu et al. 526 (154 patients); X-ray Custom MVC- Landmark: NA Mean absolute error Compared to Single clinic 10-fold patient- Custom MVC- 100 Starting:
[19]
2018 resolution: 128 × 256 Net 0.0398- (landmark): 0.0398 manual dataset; no wise cross- Net 0.01, halved
pixels 0.0459; (AP) - 0.0459 measurement and metal artifact validation; every 10
Cobb: (LAT); circular mean other deep images comparison to epochs
4.04°- absolute error (Cobb learning methods manual “gold
4.07° angle): 4.04° (AP) - standard”
4.07° (LAT)
Wang et 526; resolution: X-ray Custom MVE- Cobb: NA Circular mean Compared to Single clinic Used same Custom MVE- NA Starting:
[20]
al. 2019 0.26 mm/pixel Net 6.26°-7.81° absolute error (Cobb manual dataset dataset as Net 0.01
angle): 7.81° (AP) - measurement and previous study;
6.26° (LAT); SMAPE other deep compared to other
(Cobb angle): learning methods deep learning
24.94% (AP) - methods
11.90% (LAT)

Zhang et 2,738 pairs (AP & LAT X-ray Custom MPF- Landmark: NA Scaled mean Compared to Single clinic 10-fold cross- Custom MPF- 120 Initial:
[21]
al. 2022 X-rays); from local Net 0.0046- absolute error manual dataset validation; Net 0.001,
hospital 0.0050; (landmark): 0.0046 measurement and comparison to decayed by
Cobb: (AP) - 0.0050 other deep manual “gold 0.2 every 30
3.52°-4.05° (LAT); circular mean learning methods standard” epochs
absolute error (Cobb measurements
angle): 3.52° (AP) -
4.05° (LAT); SMAPE
(Cobb angle):
13.71% (AP) -
12.60% (LAT)
Zerouali 100 patients with X-ray SmartXpert ≤ 2.9° or ≤ ≥ 0.85 except Mean absolute error: Compared to Mainly pediatric Comparison to NA NA NA
et al. coronal & sagittal (Milvue) 2.7 mm thoracic ≤ 2.9° or ≤ 2.7 mm measurements by population, “gold standard”
[22]
2023 whole spine kyphosis = for parameters; senior exclusions manual
radiographs 0.58 intraclass correlation musculoskeletal restricted measurements;
coefficient: ≥ 0.85 radiologist analysis to visual assessment
except thoracic (ground truth) preoperative of reliability by
kyphosis = 0.58 patients radiologists
Korez et 145 images to train X-ray RetinaNet + U- 1.2°-5.0° NA Mean absolute Compared to Single center Statistical analysis RetinaNet + U- NA NA
[23]
al. 2020 model, 97 test images Net CNNs difference vs. manual manual data; did not (mean absolute Net
with variety of measurements: 1.2°- measurements by evaluate intra- difference,
conditions 5.0° for parameters spine surgeon observer correlation