Deep learning-Based Correlation Analysis of Pelvic and Spinal Sequences for Enhanced Sagittal Spinal Alignment Prediction

Kai Song, Huan Qi, Chi Ma, Pengfei Chi, Junyu Lin, Qiang Yang, Yang Cao, Bing Wang, Fangcai Li, Zezhang Zhu, Wei‐Shi Li, Jianguo Zhang, William W. Lu, Zheng WangmedRxiv 2023 · DOI

TL;DR

Researchers have developed a deep learning system that moves beyond simple geometric formulas to predict spinal alignment. By treating the pelvis and spine as a sequence of spatial key points and using a "Seq2Seq" architecture, the model significantly outperforms traditional methods in predicting critical surgical parameters like Lumbar Lordosis, especially in complex cases.

What Problem Does This Paper Address?

In spinal surgery, "sagittal alignment"—the side-view curve of your spine—is everything. To plan a successful surgery, doctors need to predict a patient’s ideal spinal curvature based on their pelvic structure.

For decades, the gold standard has been a single measurement called Pelvic Incidence (PI). The idea was simple: PI is a fixed anatomical angle, and you can use basic linear formulas to calculate what a patient's Sacral Slope (SS), Pelvic Tilt (PT), and Lumbar Lordosis (LL) should be.

However, the "PI-based" approach has major flaws:

  1. Oversimplification: It reduces complex pelvic anatomy to a single number, ignoring subtle morphological details.
  2. Lack of Specificity: It can’t predict the specific shape of local spinal segments.
  3. Invisibility: It provides a numerical target but no visual map for the surgeon.
  4. Rotation Noise: Pelvic rotation during X-rays often throws off measurements, leading to inaccurate predictions.

Methods

The authors proposed an AI-based system that treats the relationship between the pelvis and the spine as a translation task—similar to how Google Translate converts a sequence of English words into French.

The Architecture

  • Seq2Seq Framework: The team used a Sequence-to-Sequence model (typically used in Natural Language Processing) to model the spatial correlation between 25 pelvic key points and 59 spinal key points.
  • Angle Correlation Network (ACN): To solve the "rotation" problem, they introduced a novel sub-network that predicts a "neutral position." It essentially "de-rotates" the input data to minimize training loss, ensuring the model isn't confused by how a patient was standing during the X-ray.
  • Zig-Zag Offset Scheme: Instead of predicting absolute coordinates, the decoder predicts "offset vectors" in a zig-zag path (from one vertebral corner to the next). This makes it much easier for the neural network to learn the relative geometry of the spine.
  • Procrustes Analysis: Before the AI sees the data, the researchers used this statistical method to standardize the scale and orientation of all pelvic shapes, allowing the model to focus purely on morphology.

Datasets

The study utilized a high-quality, manually annotated dataset:

  • Training/Validation: 337 healthy adults (ages 18–35) with full-length lateral radiographs. 274 cases were used for training and 63 for validation.
  • Testing: A separate multi-center dataset of 51 cases was used to evaluate the model against traditional clinical formulas.
  • Ground Truth: Two expert orthopedic surgeons reached a consensus to manually annotate 84 key points per X-ray.

Key Insights and Ideas

  • From Scalars to Sequences: The biggest "aha" moment is treating anatomy as a sequence. By using 25 points for the pelvis instead of one angle (PI), the model captures the "harmony" of the entire spinopelvic complex.
  • Handling the "Extreme" Patient: Traditional formulas work okay for the "average" person but fail at the ends of the bell curve. The AI showed significantly better performance at these extreme values, which are often the cases that most need surgery.
  • Visual Planning: Because the model outputs coordinates, it can generate a predicted "ideal" spinal shape. This gives surgeons a direct visual template to aim for during reconstruction, something a simple formula cannot provide.

Key Results

The AI-based method consistently outperformed the traditional PI-based formulas across all metrics:

Metric Parameter PI-Based Method (R-value) AI-Based Method (R-value)
Correlation (R) Sacral Slope (SS) 0.67 0.80
Pelvic Tilt (PT) 0.52 0.73
Lumbar Lordosis (LL) 0.48 0.76
  • Accuracy: For Lumbar Lordosis (LL), the AI reduced the Mean Absolute Error (MAE) from 6.69 degrees down to 5.52 degrees.
  • Robustness: The AI maintained high accuracy even when the input pelvic images were randomly rotated between -60 and +60 degrees, proving the effectiveness of the Angle Correlation Network.

Limitations and Future Work

While promising, the authors acknowledge a few hurdles:

  • Manual Bottleneck: The system currently relies on surgeons manually clicking key points on X-rays. Future versions need automated landmark detection to be practical in a busy clinic.
  • 2D vs. 3D: The study used 2D lateral X-rays. The researchers aim to expand this to 3D morphological learning using CT scans or EOS imaging systems.
  • Dataset Diversity: While the multi-center test set is a good start, training on a larger, more ethnically and age-diverse population would improve the model's generalizability.

Overall, this research marks a shift from "geometric rule-of-thumb" medicine to personalized, AI-driven surgical planning.

Summary updated

Abstract

Abstract Background Pelvic Incidence (PI) plays a crucial role in surgical planning. However, it is insufficient for accurately predicting spinal alignment parameters, including Sacral Slope, Pelvic Tilt, and Lumbar Lordosis. We have devised an AI-based method for predicting sagittal spinal alignments with enhanced precision. Methods We have developed an AI-based system utilizing a Seq2Seq framework to model the spatial correlation between pelvic and spinal key points. This system was trained on a dataset of 337 cases and evaluated using 51 cases obtained from a multi-centre hospital. To address the issue of pelvic rotation, we introduced an Angle Correlation Network. We compared the performance of our AI-based system in predicting spinal alignment against the traditional PI-based method. This comparison was conducted using Mean Absolute Error (MAE) and the Correlation Coefficient (R value) as evaluation metrics. Results We evaluated the performance of our AI-based system for predicting Sacral Slope (SS), Pelvic Tilt (PT), and Lumbar Lordosis (LL) values. The Pearson correlation coefficient of the AI-based method surpassed that of the PI-based method (0.80 vs 0.67 for SS, 0.73 vs 0.52 for PT, and 0.76 vs 0.48 for LL), indicating a more robust linear relationship between AI predictions and actual values. Additionally, the AI-based method exhibited a lower Mean Absolute Error (MAE) compared to the PI-based method for LL (5.52 vs 6.69), signifying enhanced prediction accuracy. Conclusions In this study, we demonstrated the potential of an AI-based approach for predicting sagittal spinal alignments with improved precision compared to the traditional PI-based method. The AI-based system, utilizing a Seq2Seq framework and an Angle Correlation Network, exhibited a stronger linear relationship between predicted and actual values for Sacral Slope, Pelvic Tilt, and Lumbar Lordosis, as well as a reduced Mean Absolute Error for Lumbar Lordosis. These findings support the integration of AI in spinal surgery planning and personalized medicine for sagittal alignment evaluation and management.
BibTeX
@misc{song2023deep,
  author = {Kai Song and Huan Qi and Chi Ma and Pengfei Chi and Junyu Lin and Qiang Yang and Yang Cao and Bing Wang and Fangcai Li and Zezhang Zhu and Wei‐Shi Li and Jianguo Zhang and William W. Lu and Zheng Wang},
  title = {{Deep learning-Based Correlation Analysis of Pelvic and Spinal Sequences for Enhanced Sagittal Spinal Alignment Prediction}},
  year = {2023},
  doi = {10.1101/2023.09.17.23295663},
  howpublished = {medRxiv},
}
Figure