Abstract
The LanDet model exhibits competitive performance compared to existing literature. The effectiveness of the physics-informed constraint method, utilized in our landmark detection as object algorithm, is highlighted. Furthermore, we addressed the limitations of heatmap-based methods for anatomical landmark detection and tackled issues related to mis-identifying of similar or adjacent landmarks instead of intended landmark using this novel approach.
Introduction
Evaluating the sagittal balance of the spine and pelvis is critical for diagnosing spinal disorders and planning surgical interventions. Clinicians traditionally rely on manual measurements of parameters such as Sacral Slope (SS), Pelvic Tilt (PT), and the Sagittal Vertical Axis (SVA). However, manual annotation is time-consuming and prone to significant inter-observer variability.
While previous AI attempts have utilized heatmap-based regression to automate these measurements, they face three major hurdles: quantization errors that require high computational costs for precision, signal overlap between close landmarks, and anatomical confusion where models struggle to distinguish between similar-looking adjacent structures like different vertebrae.
Methods
The authors introduce LanDet (Landmark Detection), a framework that shifts the paradigm from traditional heatmap regression to an anchor-based object detection framework, similar to the YOLO architecture.
Landmarks as Objects
Instead of predicting a single pixel, LanDet treats each anatomical landmark as a detectable "object" by predicting a bounding box centered on it. This approach significantly reduces the computational overhead associated with maintaining high-resolution heatmaps while improving detection robustness.
Physics-Informed Constraints
A key innovation is the embedding of geometric physics-informed constraints directly into the model’s loss function. The authors modified the standard object detection loss to include a geometric constraint term ($l_{cnst}$), which penalizes the model if predicted landmarks violate known anatomical relationships. For example, by treating the sacrum as a solid body with predictable distances between its corners, the model can "infer" the location of obscured landmarks based on visible ones.
Datasets and Experimental Setup
The study utilized a combined dataset of 1,470 images from two distinct sources to ensure technological generalizability:
- DS1: 1,150 lateral spine X-rays from Grand River Hospital (standard X-ray devices).
- DS2: 320 lateral lumbar spine and pelvic images from Intellijoint Surgical (EOS imaging system).
The dataset included challenging clinical cases often excluded from AI studies, such as images with poor contrast, partial visibility, or existing spinal and hip implants.
Key Results
LanDet demonstrated high accuracy and surgeon-level consistency across all targeted spinopelvic measures:
- Pelvic Tilt (PT): Achieved a Mean Absolute Error (MAE) of 1.3° with a correlation ($R$) of 0.98.
- Sacral Slope (SS): Achieved an MAE of 3.7° ($R = 0.89$).
- Sagittal Vertical Axis (SVA): Achieved an MAE of 2.1 mm ($R = 0.96$).
- Reliability: The model achieved an Intraclass Correlation Coefficient (ICC) > 0.88 compared against three senior surgeons.
- Robustness: The system successfully detected landmarks in 98% of test images, including those featuring significant metal hardware that typically causes automated tools to fail.
Conclusion
By incorporating physics-informed constraints and treating landmarks as objects, LanDet provides a holistic understanding of spinal anatomy that overcomes the "adjacent landmark trap." While future work will focus on larger, more diverse datasets and 3D expansion for CT/MRI data, LanDet currently offers a highly efficient and robust solution for automated clinical spinopelvic assessment.
Summary updated