3. Automated Lenke classification for preoperative spine surgery by extracting anatomical landmarks from X-ray images using a deep learning approach

BACKGROUND CONTEXT

Spinopelvic assessment (eg, SS, PT, PI, LL, TK, CL, SVA, and Cobb angle) is vital for preoperative spinal surgery planning but is often measured manually, leading to variability. Recent AI and deep learning methods improve automation and accuracy. While promising, these techniques face challenges including computational complexity, small test datasets, lack of surgeon validation, and limited robustness to varied image conditions.

PURPOSE

To increase accuracy, reduce complexity, and provide robust preoperative X-ray analysis, we propose a novel, physics-informed deep learning method based on mathematical spinal relations. This approach aims to automatically calculate lateral and AP spinal parameters and promptly perform Lenke classification for each patient.

STUDY DESIGN/SETTING

N/A

PATIENT SAMPLE

We collected 3500 lateral and AP spine X-rays from Grand River Hospital (GRH) in Kitchener, ON, Canada, between 2016 and 2024, encompassing hip/spine implants, varied postures, and poor-contrast or partially visible spines. Image processing filters enhanced annotation accuracy, allowing landmark detection even in incomplete images. The dataset includes conventional and EOS systems, enabling thorough performance evaluation and robust landmark detection. Data was split into 80% training, 10% validation, and 10% testing.

OUTCOME MEASURES

This study focuses on the automatic extraction of spinopelvic parameters and anatomical landmarks from lateral and AP X-ray images, including SS, PT, PI, LL, SVA, femur center, sacrum end plate, iliac crest, L1–L5, T12–T1, C7–C2, apex, Cobb angle, LSRS, TSM, and CSRS. These measurements enable Lenke classification, identifying curve types (1–6), lumbar modifiers (A, B, C), and thoracic modifiers (–, N, +). To evaluate performance, we use relative root mean square error (RRMSE) to compare predicted values (PR) with manual annotations (MA), while intraclass correlation coefficient (ICC) measures reliability among surgeons, MA, and PR.

METHODS

Using our developed physics-informed deep learning method, spinopelvic parameters were extracted from X-ray images and validated against manual annotations. Landmarks were detected as objects with geometric constraints derived from mathematical spinal relations. Performance, compared to three senior spine surgeons, demonstrated excellent correlation, with intraclass correlation coefficients exceeding 0.9, surpassing previously reported literature values. Additionally, we developed an algorithm leveraging these parameters to automate Lenke classification, identifying curve type (1–6), lumbar modifier (A,B,C), and thoracic modifier (–,N,+), significantly aiding triage and preoperative planning.

RESULTS

We evaluated our model on the dataset, achieving final accuracies of 93.1% (SS), 94.6% (PT), 93.4% (Cobb angle), 91.2% (LL), and 94.5% (SVA). Patient classification attained 98.5% accuracy via our automated Lenke-based algorithm. Overall, the model surpasses literature-reported accuracy, demonstrating robust performance and reliability. To compare against surgeons, we used the intraclass correlation coefficient (ICC) with three surgeons’ annotations, revealing stronger consistency than previously reported.

CONCLUSIONS

Our physics-informed deep learning method reliably automates spinopelvic parameter extraction and classification, achieving high accuracy and robust surgeon-level consistency, thus advancing preoperative spinal planning and guiding AI innovations.

FDA Device/Drug Status

This abstract does not discuss or include any applicable devices or drugs.