Ensemble learning of deep CNN models and two stage level prediction of Cobb angle on surface topography in adolescents with idiopathic scoliosis.
Authors
Affiliations (5)
Affiliations (5)
- Department of Mechanical Engineering, University of Alberta, 10th Floor, Donadeo Innovation Centre for Engineering, Edmonton, T6G 1H9, AB, Canada; Mechanical Design & Production Department, Faculty of Engineering, Cairo University, Cairo University Rd, Oula, Giza District, Cairo, 12613, Egypt. Electronic address: [email protected].
- Department of Mechanical Engineering, University of Alberta, 10th Floor, Donadeo Innovation Centre for Engineering, Edmonton, T6G 1H9, AB, Canada.
- Allied Health Institute, Federal University of Mato Grosso do Sul, Cidade Universitária, Av. Costa e Silva, Campo Grande, 79070-900, MS, Brazil.
- Department of Civil and Environmental Engineering, University of Alberta, 7th Floor, Donadeo Innovation Centre for Engineering, Edmonton, T6G 1H9, AB, Canada.
- Department of Mechanical Engineering, University of Alberta, 10th Floor, Donadeo Innovation Centre for Engineering, Edmonton, T6G 1H9, AB, Canada; Department of Biomedical Engineering, University of Alberta, 13th Floor, Donadeo Innovation Centre for Engineering, Edmonton, T6G 1H9, AB, Canada.
Abstract
This study employs Convolutional Neural Networks (CNNs) as feature extractors with appended regression layers for the non-invasive prediction of Cobb Angle (CA) from Surface Topography (ST) scans in adolescents with Idiopathic Scoliosis (AIS). The aim is to minimize radiation exposure during critical growth periods by offering a reliable, non-invasive assessment tool. The efficacy of various CNN-based feature extractors-DenseNet121, EfficientNetB0, ResNet18, SqueezeNet, and a modified U-Net-was evaluated on a dataset of 654 ST scans using a regression analysis framework for accurate CA prediction. The dataset comprised 590 training and 64 testing scans. Performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and accuracy in classifying scoliosis severity (mild, moderate, severe) based on CA measurements. The EfficientNetB0 feature extractor outperformed other models, demonstrating strong performance on the training set (R=0.96, R=20.93) and achieving an MAE of 6.13<sup>∘</sup> and RMSE of 7.5<sup>∘</sup> on the test set. In terms of scoliosis severity classification, it achieved high precision (84.62%) and specificity (95.65% for mild cases and 82.98% for severe cases), highlighting its clinical applicability in AIS management. The regression-based approach using the EfficientNetB0 as a feature extractor presents a significant advancement for accurately determining CA from ST scans, offering a promising tool for improving scoliosis severity categorization and management in adolescents.