Abstract
Automated scoliosis assessment relies on the accurate identification of vertebral landmarks in X-ray images. While deep learning models have shown promise, standard training objectives like Mean Squared Error (MSE) often fail to account for the inherent geometric structure of the spine, leading to anatomically inconsistent predictions. This paper presents a novel Bipartite Distance (BPD) loss function that models the spine as a bipartite graph. By enforcing pairwise distance constraints between left and right landmarks, the model learns to respect the global spinal shape. Experimental results demonstrate that the BPD-enhanced model achieves higher correlation with ground truth and produces more realistic anatomical curves than baseline methods.
Introduction
Scoliosis is characterized by a lateral curvature of the spine, requiring precise measurement for effective diagnosis and treatment. Clinicians typically determine the degree of curvature by manually identifying the corners of vertebrae in anterior-posterior X-ray images. However, this manual process is labor-intensive and suffers from significant inter-observer variability.
Convolutional Neural Networks (CNNs) offer a path toward automation. Yet, most landmark detection models utilize MSE as the primary loss function. MSE treats each landmark as an independent coordinate, ignoring the rigid biological relationships between them. This lack of structural awareness often results in "jittery" or physically impossible spinal reconstructions. To address this, we propose integrating a geometric constraint directly into the learning process.
Methods: The Bipartite Distance Approach
The proposed method moves beyond point-wise regression by incorporating a Bipartite Distance (BPD) loss. The spinal landmarks are partitioned into two sets: Set A (left-side landmarks) and Set B (right-side landmarks). The spine is then modeled as a complete bipartite graph where every node in Set A is connected to every node in Set B.
The BPD loss function calculates the Euclidean distances between these connected pairs in both the ground-truth data and the model’s predictions. By minimizing the difference between these distances, the model is forced to maintain the "width" and "shape" of the spine. The final training objective is a weighted combination: $L = MSE + \alpha BPD$, where $\alpha$ is a hyperparameter (set to 0.01 in this study) that balances local coordinate accuracy with global structural coherence.
Experimental Setup
The study utilized a dataset of 100 high-resolution anterior-posterior spinal X-ray images. The model targets 18 vertebrae (C7, T1–T12, and L1–L5), with 4 corners per vertebra, totaling 72 landmarks per image. Images were resized to $1024 \times 512$ and normalized before being processed by an end-to-end CNN architecture. The data was split into training (80), testing (15), and validation (5) sets.
Key Results
The inclusion of BPD loss significantly improved both the quantitative accuracy and the qualitative realism of the predicted landmarks.
Quantitative Performance
The BPD-enhanced model achieved a Pearson correlation coefficient of 0.95, a notable improvement over the 0.92 achieved by the baseline MSE-only model. Furthermore, a one-way ANOVA analysis indicated no significant difference between the landmarks predicted by the BPD model and those provided by medical experts, validating its clinical potential.
Qualitative Improvements
Visually, the difference between the two models is stark. As shown in the comparison below, the standard MSE model often produces irregular, jagged shapes where vertebrae appear distorted. In contrast, the MSE-BPD model generates smooth, anatomically consistent curves that closely follow the natural structure of the spine.
Discussion and Conclusion
The core contribution of this work is the shift from treating medical landmark detection as a simple regression task to treating it as a structural problem. By enforcing global geometric priors through the BPD loss, the model gains an implicit understanding of spinal anatomy without requiring complex multi-stage pipelines or manual feature engineering.
While the results are promising, the study is limited by its small dataset size and its focus on 2D projections of a 3D deformity. Future work will explore extending the bipartite distance logic to 3D imaging modalities and testing the model's robustness on larger, more diverse clinical datasets.
Summary updated