Abstract
Accurate assessment of scoliosis is critical for clinical decision-making, yet traditional manual measurement of the Cobb angle is prone to significant inter-observer variability. This paper introduces the Adaptive Error Correction Net (AEC-Net), a deep learning framework designed to automate spine curve analysis with high precision. By integrating landmark-based detection with direct angle regression and a novel numerical error-correction layer, AEC-Net achieves a mean absolute error (MAE) of 4.90°. This performance represents a substantial improvement over existing automated methods, offering a robust tool for clinical-grade scoliosis diagnosis.
Introduction
Scoliosis diagnosis and monitoring rely heavily on the Cobb angle, the gold standard metric calculated from spinal X-rays. In clinical practice, radiologists manually identify the most tilted vertebrae in a curve and measure the angle between them. This process is not only time-consuming but also suffers from high variability between different experts.
Automating this measurement is challenging due to the low contrast of X-ray images and the anatomical ambiguity of vertebral boundaries. Previous computational approaches generally followed one of two paths: landmark-based methods, which are sensitive to minor coordinate errors, or direct estimation methods, which often lack the fine-grained anatomical detail required for precision. AEC-Net addresses these limitations by combining both approaches through an adaptive correction mechanism.
Methodology
The AEC-Net architecture is composed of three specialized sub-networks designed to capture complementary features of spinal geometry:
- Landmark Net: Focuses on spinal boundary features to regress 68 specific landmarks, providing an indirect Cobb angle calculation ($angle'$).
- Angle Net: Focuses on the global spinal curve to regress Cobb angles directly from raw image pixels ($angle''$).
- Adaptive Error Correction Net: The core innovation of the system. It utilizes extrapolation—a classical numerical mathematics technique—to adaptively offset the errors inherent in the first two networks.
The framework employs a two-stage Iterative Angle Training (IAT) process. This optimization strategy leverages the "error complementarity" between the landmark-based and direct-estimation outputs, allowing the correction net to learn and mitigate systematic biases.
Experimental Setup and Datasets
The study utilized a dataset of 581 spinal anterior-posterior X-ray images from subjects with varying degrees of scoliosis. The analysis targeted 17 vertebrae across the thoracic and lumbar spine, involving 68 landmark points and three distinct Cobb angles (Proximal Thoracic, Main Thoracic, and Thoracolumbar/Lumbar) per image. Ground truth annotations were provided by two professional experts from a specialized spine center. The dataset was split into 481 images for training and validation, with 100 unseen images reserved for final testing.
Results and Discussion
AEC-Net demonstrated superior performance compared to traditional benchmarks and its own sub-components:
- High Accuracy: The system achieved an overall MAE of 4.90°. Notably, the error correction step effectively halved the error of the best individual sub-network (Angle Net at 9.36° and Landmark Net at 13.85°).
- Clinical Consistency: The system showed high correlation with expert measurements, with Pearson coefficients ranging from 0.903 to 0.945 across different spinal sections.
- Robustness: In tests involving simulated Gaussian noise, AEC-Net maintained a lower Relative Root Mean Squared Error (RRMSE) than standard methods like Support Vector Regression (SVR), indicating its suitability for varying image qualities found in clinical settings.
Conclusion
AEC-Net provides a highly accurate and automated solution for Cobb angle measurement, bridging the gap between boundary-focused landmark detection and global curve estimation. While future work will focus on expanding dataset diversity and exploring 3D reconstructions, the current 2D system offers a significant step toward reducing clinical workload and improving the consistency of scoliosis assessments.
Summary updated