You are here
Home > Blog > Family Practice Medicine > Scoliosis In Adolescents Identified By Comprehensive AI Technology

Scoliosis In Adolescents Identified By Comprehensive AI Technology

Scoliosis In Adolescents Identified By Comprehensive AI Technology

Adolescent idiopathic scoliosis (AIS) is one of the most prevalent spinal deformities, impacting millions globally. The Cobb method has been the cornerstone for measuring spinal curvature and guiding treatment decisions. However, its application is hindered by significant inter-observer variability and time-consuming procedures. Emerging technologies, including smartphone-aided measurement and artificial intelligence (AI), have begun to address these challenges. This study introduces a groundbreaking automated AIS measurement system utilizing convolutional neural networks to accurately identify vertebral boundaries and classify spinal curvature according to the Lenke system. This innovative approach promises to enhance precision, reliability, and efficiency in AIS diagnosis and treatment planning, significantly reducing clinician workload.

 

THE STUDY BACKGROUND

Adolescent idiopathic scoliosis (AIS) is a prevalent spinal deformity affecting millions of children and adolescents globally [1]. The Cobb method has long been the gold standard for calculating spinal curvature angles, significantly influencing AIS patients’ diagnosis, treatment planning, and surgical decision-making processes [2, 3]. However, despite its widespread use, the Cobb method is plagued by notable limitations. Interobserver variability, often ranging from 4° to 8°, arises from differences in identifying vertebral endplates, and the measurement process is time-consuming, requiring 15 to 20 minutes per patient.

Recent technological advancements have sought to mitigate these challenges. New methods, such as the end-vertebra tilt angle technique and smartphone-aided measurement tools, have streamlined the measurement process, reducing the reliance on manual techniques and improving accuracy [4-6]. Moreover, the integration of image processing technology and artificial intelligence (AI) has paved the way for fully automated spinal shape analysis. Deep learning models, which label vertebral endplate centers, have shown promise in automating the measurement of Cobb angles. However, they primarily focus on standing coronal views. They cannot accurately identify cephalad and caudad borders in lateral and side-bending views [7-9].

The limitations of current automated methods highlight the need for more advanced solutions that can meet clinical demands comprehensively. The inability to automatically identify the cephalad and caudad borders within their respective ranges poses a significant barrier to the widespread clinical adoption of these technologies [10, 11]. In response to these challenges, this study proposes developing an automatic AIS measurement system utilizing convolutional neural networks. This system aims to automatically determine spinal curves, identify vertebral borders, measure the Cobb angle, and classify the curvature according to the Lenke classification. 

The study seeks to validate this AI-based measurement system’s feasibility, accuracy, and reliability by comparing its performance with manual measurements conducted by experienced spinal surgeons. By addressing the existing limitations in AIS measurement and classification, this innovative system has the potential to enhance precision, reduce clinician workload, and improve overall treatment outcomes for AIS patients.

 

THE STUDY METHOD

Data Collection

This retrospective study, approved by the Ethics Committee of Guizhou Orthopedic Hospital in July 2021, involved the collection of X-ray images from 530 patients. After excluding 30 cases with congenital scoliosis, neuromuscular disorders, and syndromic disorders, 500 images remained. These images, desensitized to protect patient identity, were randomly divided into a training set of 300 photos and a validation set of 200 images. The training set was used to train the automatic measurement systems. In contrast, the validation set compared the accuracy of these systems against manual measurements.

Construction of the Automatic Measurement System

The automatic measurement system was built using three deep convolutional network architectures: Residual Network (ResNet), DenseNet, and Inception Network. The ResNet model, consisting of eight convolutional layers, four max-pooling layers, and one fully connected layer, was optimized to handle the gradient vanishing problem. DenseNet, known for its generalization capabilities, incorporated dense blocks and transition layers to mitigate overfitting. The Inception Network utilized multiple small convolutional kernels to enhance parameter usage and computation speed. Each model was meticulously designed with specific configurations of convolutional kernels, pooling sizes, and activation functions to achieve optimal performance. The training set’s X-ray images were labeled with color-coded points marking the vertices of each vertebra from T1 to L5, which were then used to train the deep learning algorithms for vertebral segmentation, key point detection, and Cobb angle calculation.

Oscillogram and Lenke Classification

The study introduced oscillograms to visualize the spinal curvature and facilitate the classification according to the Lenke system. Oscillograms, derived from the tilt angles of vertebral endplates, showcased unique patterns corresponding to different Lenke types. These patterns allowed for the identification of structural bends that are essential for accurate Cobb angle measurement and classification. The A.I. system utilized these oscillograms to distinguish between various curve types and structural characteristics, aiding in the diagnosis and treatment planning for AIS patients.

Measurement Methods and Evaluation

The study evaluated the A.I. system’s accuracy by comparing its measurements against those obtained manually by senior and junior spine surgeons using reliable software for spinal measurement, Surgimap. According to the Lenke system from the imported radiographs, the A.I. system automatically outputs the Cobb angles, upper and lower vertebral locations, and classifications. The study assessed multiple measurement indicators, including the Cobb angles of various spinal segments, thoracic sagittal profile, bending views, and sagittal thoracic alignment. Image quality was also evaluated to determine its influence on the accuracy of critical point identification, with assessments conducted on bone/soft tissue contrast, bone sharpness, and other relevant dimensions.

 

ANALYSIS

The researcher conducted the statistical analysis using SPSS software (version 25.0, IBM). Comparative studies between groups were performed using t-tests for continuous variables and χ²-tests for categorical variables. To assess manual evaluators’ reliability, intraobserver and interobserver reliabilities were estimated by calculating the intraclass correlation coefficient (ICC) with corresponding 95% confidence intervals (CI). Furthermore, the study evaluated the consistency between the Cobb angles derived from the automatic measurement system and manual methods using the mean absolute error (MAE) and ICC.

 

RESULTS

Demographic Data

– Patients: The study included 500 patients aged 8 to 24 years (average age 13.55 years), with 86 males and 414 females. Among them, 200 patients were part of the validation set.

– Curve Characteristics: Measurements ranged from 0.0 to 59.9 degrees for proximal thoracic (PT), 7.7 to 85.9 degrees for main thoracic (MT), 1.3 to 71.9 degrees for thoracolumbar/lumbar (TL/L), and 0.7 to 59.9 degrees for lateral T5-T12. The distribution of Lenke classifications included varying numbers across types 1 to 6.

 

Measuring Time Comparison

– Artificial Intelligence (A.I.) System: The A.I. system took about 200 milliseconds per patient, totaling 30 minutes for 300 patients in the training set.

– Manual Methods: Senior surgeons spent around 23.6 minutes per patient, totaling about 3600 minutes for 300 patients. It included reliability checks, which doubled to about 7100 minutes.

Consistency between A.I. and Manual Methods

– Accuracy: Comparing A.I. with manual methods (senior and junior surgeons) showed strong agreement (intraclass correlation coefficient, ICC > 0.80) and low error (mean absolute error, MAE < 7.36 degrees) for PT, MT, TL/L, and sagittal thoracic T5-T12.

– Lenke Classification: The A.I. system’s classification consistency (ICC > 0.93) indicates reliable performance across different spinal curve types.

Image Quality Assessment

– Impact on Accuracy: Images with more significant errors (> 5 degrees MAE) had lower quality scores, particularly in bone contrast and sharpness.

These findings demonstrate the effectiveness of the A.I. system in automating spinal curvature measurement, offering comparable accuracy to manual methods while significantly reducing measurement time and maintaining high reliability.

 

DISCUSSION

The study explored the application of convolutional neural networks (CNNs) for automated measurement of the Cobb angle in adolescent idiopathic scoliosis (AIS) patients. CNNs employed an encoder-decoder framework to identify vertebral boundaries and generate oscillograms, enabling rapid and simultaneous measurement of multiple spinal parameters, including the proximal thoracic (PT), main thoracic (MT), and thoracolumbar/lumbar (TL/L) curves, as well as sagittal alignments and bending views, all within an impressive 200 milliseconds per patient. This efficiency contrasts starkly with traditional manual methods, which consumed approximately 7100 minutes for 300 patients, illustrating the significant time-saving potential of A.I. in clinical practice [1].

A pivotal innovation in this approach was using oscillograms derived from endplate slope distributions to facilitate Lenke classification—a critical factor in scoliosis treatment planning. Previous studies have shown high agreement between automatic and manual methods in Cobb angle measurements, underscoring the reliability of AI-driven techniques in spinal curvature assessment [2]. This method extended previous work by automating the identification of vertebral boundaries based on 68 manually labeled vertebral points and subsequent training with 200 X-ray images from AIS patients. It also expedited measurements and enhanced consistency by minimizing human error associated with manual annotation [3]. 

Despite the success in PT and MT measurements, challenges persisted in accurately capturing sagittal thoracic alignments due to occlusions from thoracic, scapular, and humeral structures. Experts expect future advancements in 3D recognition technology to overcome these limitations and improve overall accuracy in complex spinal assessments [4].

The introduction of oscillography to delineate structural bends represents a notable advancement in spinal diagnostics. By transforming complex radiographic data into computationally digestible oscillograms, this method provided a standardized approach to evaluate scoliosis severity, guiding tailored treatment decisions. This automated approach not only matched but sometimes exceeded the diagnostic precision of junior spine surgeons, particularly in Lenke classification, thereby potentially enhancing clinical outcomes through more informed surgical planning [5].

In conclusion, while further refinement and larger-scale data training are needed to optimize the reliability of oscillogram-based A.I. systems, this study demonstrates substantial progress toward enhancing the efficiency and accuracy of spinal curvature assessments. The integration of A.I. technologies promises to revolutionize the management of AIS by offering faster, more consistent diagnostic capabilities, ultimately improving patient care and treatment outcomes in orthopedic practice.

 

STUDY LIMITATIONS

  1. Variability in Imaging Data

   – Imaging data from different medical institutions may vary in clarity, projection angle, and stitching.

   – Though seemingly minor to human physicians, these variations can create significant recognition gaps for the computer.

  1. Single-Center Training Dataset

   – The model was trained using a dataset collected from a single center.

   – This limitation restricts the model’s broader applicability, highlighting the need for training with data from multiple centers to improve generalizability.

 

  1. Image Quality Differences

   – Differences in image quality can affect the computer’s ability to recognize and process the data accurately.

   – Ensuring consistent image quality across different institutions is essential for reliable A.I. performance.

  1. Projection Angle and Stitching Issues

   – Inconsistent projection angles and image stitching can introduce errors in automated measurements.

   – Standardizing imaging protocols can help mitigate these issues and improve the accuracy of AI-based assessments.

  1. Necessity for Larger-Scale Data

   – Expanding the training dataset to include a more diverse range of images from multiple centers is crucial.

   – Larger-scale data training will enhance the reliability and robustness of the A.I. system in various clinical settings.

By addressing these limitations, the study aims to reduce the burden on surgeons by automating repetitive tasks, thereby improving measurement accuracy and the feasibility of A.I. methods in clinical practice.

 

CONCLUSION

The A.I. system demonstrates high reliability in performing the Lenke classification for adolescent idiopathic scoliosis, suggesting it could be a valuable auxiliary tool for spinal surgeons. By automating this critical assessment, the A.I. system can save significant time and reduce the workload for surgeons, allowing them to focus on more complex and nuanced aspects of patient care. This advancement holds promise for enhancing scoliosis diagnosis and treatment planning efficiency and accuracy.

 

References

  1. Weinstein SL, Dolan LA, Cheng JC, Danielsson A, Morcuende JA. Adolescent idiopathic scoliosis. Lancet. 2008;371(9623):1527–37. [https://doi.org/10.1016/S0140-6736(08)60658-3

 

  1. Cobb JR. Scoliosis; quo vadis. J Bone Jt Surg Am. 1958;40:507–10. [https://journals.lww.com/jbjsjournal/Citation/1958/40030/Scoliosis__Quo_Vadis_.14.aspx]

 

  1. Altaf F, Gibson A, Dannawi Z, Noordeen H. Adolescent idiopathic scoliosis. BMJ. 2013;346:f2508. [https://doi.org/10.1136/bmj.f2508]

 

  1. Qiao J, Liu Z, Xu L, Wu T, Zheng X, Zhu Z, et al. Reliability analysis of a smartphone-aided measurement method for the Cobb angle of scoliosis. J Spinal Disord Tech. 2012;25(4):E88–92. [https://doi.org/10.1097/BSD.0b013e3182463964]

 

  1. Wang J, Zhang J, Xu R, Chen TG, Zhou KS, Zhang HH. Measurement of scoliosis Cobb angle by end vertebra tilt angle method. J Orthop Surg Res. 2018;13(1):223. [https://doi.org/10.1186/s13018-018-0928-5]

 

  1. Shaw M, Adam CJ, Izatt MT, Licina P, Askin GN. Use of the iPhone for Cobb angle measurement in scoliosis. Eur Spine J. 2012;21(6):1062–8. [https://doi.org/10.1007/s00586-011-2059-0]

 

  1. Zhang J, Lou E, Le LH, Hill DL, Raso JV, Wang Y. Automatic Cobb measurement of scoliosis based on fuzzy Hough transform with vertebral shape prior. J Digit Imaging. 2009;22(5):463–72. [https://doi.org/10.1007/s10278-008-9127-y]

 

  1. Wu H, Bailey C, Rasoulinejad P, Li S. Automated comprehensive adolescent idiopathic scoliosis assessment using MVC-net. Med Image Anal. 2018;48:1–11. [https://doi.org/10.1016/j.media.2018.05.005]

 

  1. Galbusera F, Niemeyer F, Wilke HJ, Bassani T, Casaroli G, Anania C, et al. Fully automated radiological analysis of spinal disorders and deformities: a deep learning approach. Eur Spine J. 2019;28(5):951–60. [https://doi.org/10.1007/s00586-019-05944-z]

 

  1. Sun Y, Xing Y, Zhao Z, Meng X, Xu G, Hai Y. Comparison of manual versus automated measurement of Cobb angle in idiopathic scoliosis based on a deep learning keypoint detection technology. Eur Spine J. 2022;31(8):1969–78. [https://doi.org/10.1007/s00586-021-07025-6]

 

  1. Huang X, Luo M, Liu L, Wu D, You X, Deng Z, et al. The comparison of convolutional neural networks and the manual measurement of Cobb angle in adolescent idiopathic scoliosis. Global Spine J. 2024;14(1):159–68. [https://doi.org/10.1177/21925682221098672]

Oncology Related Tools


Other


Latest Research


Scoliosis


About Author

Similar Articles

Leave a Reply


thpxl