Deep Learning Based Image Segmentation for Diagnosis of Spondylolisthesis from Lumbar Spine X-RAY

Abstract

Spondylolisthesis is a condition of Spine which occurs due to slippage of vertebrae. which can causepain and resultsin limited mobility and nerve compression. Accurate detection in early stages important for prevention of further complications. Traditional methos heavily depends on human examination of X-ray images, which can cause manual errors. To improve the detection of spondylolisthesis in X-ray images we used deep learning techniques. YOLO (you only look once) is known for itsreal time processing capability and high precision, works faster, detects and segments in one step based on percentage of vertebral slippage, and gives more accurate results. We compared the results obtained by YOLOV8 and YOLOV11. This study supports the use of deep learning techniques to assist medical professionals in making more accurate assessments.

Introduction

Overview:

Spondylolisthesis, commonly affecting the lumbar spine, is the forward slippage of a vertebra and is graded using the Meyerding classification. Traditional diagnosis using X-rays, MRIs, and CT scans is manual and error-prone. This study introduces an AI-driven system using YOLOv8 segmentation to automatically detect, classify, and grade spondylolisthesis from lumbar spine X-ray images. The system aims to improve speed, precision, and reliability in spinal diagnostics.

Literature Review Highlights:

• [1] Deep Transfer Learning (Fraiwan et al., 2022): Achieved up to 96.73% accuracy using CNNs for classifying spinal disorders, but suffered from dataset size and interpretability issues.

• [2] Automated CT-Based Diagnosis (Liao et al., 2016): Used machine learning on CT images for precise slippage quantification. Required high-quality annotated data and computational resources.

• [3] SO-YOLO for WBC Detection: Introduced an improved YOLO variant with microscopy, emphasizing high detection efficiency but faced challenges like high equipment costs and model transparency.

Dataset:

• BUU-LSPINE Dataset from Burapha University and Korea Institute of Oriental Medicine.

• Contains X-rays of 400 patients (AP and LA views), aged 6 to 89.

• Includes JPEG images with labels converted to YOLO format for training.

Methodology:

• Data Preparation: Resizing to 640×640 pixels; conversion of labels to YOLO format.

• Model Training: YOLOv8 used with pre-trained weights; trained using SGD optimizer. Evaluation metrics include precision, recall, mAP, and F1-score.

• Segmentation: YOLOv8 used for pixel-level segmentation to accurately highlight vertebrae.

• Data Augmentation: Techniques like histogram equalization, contrast/brightness scaling, noise addition, and gamma correction were applied to improve generalization.

• Slippage & Grading: Slippage (%) = (Displacement / Width of inferior vertebra) × 100. Severity graded using Meyerding scale (Grades I–V).

Results:

• YOLOv8 and YOLOv11 both showed perfect classification for vertebral levels (L1–S1) with 100% accuracy.

• mAP@0.5: 0.829 and F1-score: 0.83 indicate strong model performance.

• Confusion matrices show zero misclassifications, but this may indicate potential overfitting.

• Visual outputs confirmed successful segmentation and slippage detection (e.g., 7.78% slippage = Grade I).

Conclusion

This research focused on developing an automated system for detecting spondylolisthesis using deep learning models, specifically YOLOv8 and YOLOv11. The study involved training, validating, and testing medical images while incorporating advanced image processing techniques toenhance model performance. The results demonstrated that although YOLOv11 introduced improvements in architecture and feature extraction, it outperform YOLOv8 in this specific application. YOLOv11 proved to be a more suitable model for spondylolisthesis detection, offering an optimal balance between computational efficiency, accuracy, and speed. The findings validate its effectiveness in medical image analysis and reinforce the potential of deep learning in diagnostic applications. This research highlights the importance of selecting models that align with clinical needs and paves the way for further advancements in AI-driven medical imaging solutions. Whilethisstudyhasyieldedpromising results,thereareseveralareaswherefurtherresearchand improvementscanbemade:Enhancement of Model Performance – Future work can focus on refining the YOLOv11 model by optimizing hyperparameters, implementing advancedtrainingstrategies, andexploring alternativelossfunctionsto improvedetectionaccuracy.Additionally, experimentingwithhybrid models, such as Vision Transformers (ViTs), may further enhance feature extraction capabilities.Expansion and Augmentation of the Dataset – Increasing the dataset with high-resolution images from diverse sources can help improve model generalization across various patient demographics. The use of data augmentation techniques, including synthetic image generation through generative adversarial networks (GANs), may enhance model robustness and adaptability to real-world medical imaging conditions.Integrationof Multi-Modal Imaging – Incorporating additional imaging modalities, such as MRIandCT scans, alongside X-rayimages couldlead toa more comprehensive assessment of spinal abnormalities. A multi-modal approach may improve diagnostic accuracy by leveraging complementary information from different imaging techniques.Clinical ImplementationandReal-WorldDeployment– Developing acloud-basedormobileapplicationforreal-timespondylolisthesis detection can improve accessibility for healthcare professionals. The integration of the model into a Clinical Decision Support System (CDSS) could assist radiologists by providing AI-driven second opinions, ultimately enhancing diagnostic efficiency and reducing workload.

References

[1] P.Klinwichit,W.Yookwan,S.Limchareon,K.Chinnasarn,J.-S.Jang,andA.Onuean,‘‘BUULSPINE:AThaiopenlumbarspinedatasetfor spondylolisthesis detection,’’ Appl. Sci., vol. 13, no. 15, p. 8646, Jul. 2023, doi: 10.3390/app13158646. [2] M.Fraiwan,Z.Audat,L.Fraiwan,andT.Manasreh,‘‘UsingdeeptransferlearningtodetectscoliosisandspondylolisthesisfromXrayimages,’’ PLoS ONE, vol. 17, no. 5, May 2022, Art. no. e0267851. [3] Nillmani,N.Sharma,L.Saba,N.Khanna,M.Kalra,M.Fouda,andJ.Suri,‘‘Segmentationbasedclassificationdeeplearningmodelembeddedwith explainable AI for COVID-19 detection in chest X-ray scans,’’ Diagnos tics, vol. 12, no. 9, p. 2132, Sep. 2022. [4] A.Bochkovskiy,C.-Y.Wang,and H.-Y.M.Liao,‘‘YOLOv4:Optimalspeed and accuracy ofobjectdetection,’’2020,arXiv:2004.10934. [5] E.Seeram,DigitalRadiography:PhysicalPrinciplesandQualityControl.Cham,Switzerland:Springer,2019. [6] T.S.Curry,J.E.Dowdey,andR.C.Murry,Christensen’sPhysicsofDiagnosticRadiology.Philadelphia,PA,USA:LippincottWilliams& Wilkins, 1990. [7] J.R.Balland T.Price,Chesneys’RadiographicImaging.Hoboken,NJ,USA:Wiley,1995. [8] J.T.Bushberg,J.A.Seibert,E.M.Leidholdt,andJ.M.Boone,TheEssentialPhysicsofMedicalImaging.Philadelphia,PA,USA:Lippincott Williams & Wilkins, 2011. [9] E.CarverandB.Carver,MedicalImaging:Techniques,ReflectionandEvaluation.London,U.K.:ChurchillLivingstone,2012. [10] S.Zhao,X.Wu,B.Chen,andS.Li,‘‘AutomaticspondylolisthesisgradingfromMRIsacrossmodalitiesusingfasteradversarialrecognitionnet work,’’ Med. Image Anal., vol. 58, Dec. 2019, Art. no. 101533.

Copyright

Copyright © 2025 D. Mahita, A. Sravya, Ch. Shireesha, M. Jhansi, Dr. P. M. K. Prasad. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.