A Multi-Model Deep Learning Framework for Automated Lumbar Disease Diagnosis

Abstract

To improve patient outcomes and lessen the long-term cost on healthcare systems, lumbar illnesses must be accurately diagnosed and classified. Conventional diagnostic methods rely significantly on the knowledge of radiologists, which may result in irregularities and treatment delays. This article presents a novel multi-model deep learning framework that combines the advantages of several CNN architectures, such as MobileNet, DenseNet, ResNet50, and AlexNet, with a CNN-SVM hybrid and an Involutional VGG model, in order to handle this difficulty. The ability to extract and categorize intricate and delicate lumbar spine features from medical images is much improved by this ensemble technique. The diagnosis process is further made more transparent and reliable for doctors by including Gradient-weighted Class Activation Mapping (GRAD-CAM), which offers visual interpretability. A large, annotated dataset of lumbar MRI/X-ray images was used to test the suggested framework. The experimental findings demonstrate better accuracy, recall, precision, and AUC_ROC compared to single-model CNNS. This approach shows promise as a trustworthy decision-support tool for the diagnosis and tracking of lumbar illness in real time.

Introduction

Lumbar spine disorders—including spinal stenosis, herniated discs, degenerative disc disease, and spondylolisthesis—are major causes of chronic pain and disability worldwide. Accurate diagnosis, typically based on manual interpretation of X-rays or MRI scans, is time-consuming and prone to human error. Advances in deep learning, particularly Convolutional Neural Networks (CNNs), have improved automated medical image analysis by learning hierarchical features directly from raw data.

This study proposes a multi-model deep learning framework combining several architectures (MobileNet, DenseNet, ResNet50, AlexNet, VGG16 with involutional layers, and a hybrid CNN-SVM model) to enhance lumbar disease classification. The involutional layers improve spatial adaptability, and Grad-CAM visualization adds interpretability to build clinician trust.

The dataset includes annotated lumbar MRI and X-ray images covering multiple disease categories. Preprocessing involved resizing, normalization, and data augmentation. Models were trained and evaluated on metrics such as accuracy, precision, recall, and F1-score.

Results show:

• ResNet50 and the hybrid CNN-SVM model achieved perfect accuracy (100%) in classifying all lumbar conditions.

• VGG16 with involutional layers improved spatial understanding and classification accuracy.

• MobileNet and AlexNet performed well but less accurately than deeper models.

• DenseNet struggled, especially with subtle disease distinctions.

• Grad-CAM visualizations confirmed that successful models focused on key lumbar structures.

The hybrid CNN-SVM model excelled by combining CNN feature extraction and SVM classification, effectively handling complex variations and small datasets. Overall, this multi-model approach provides a reliable, efficient, and interpretable solution for automated lumbar spine disorder diagnosis.

Conclusion

A thorough deep learning -based system for the automated classification of lumbar spine disorders from MRI images is presented in this work. We show how well deep learning works to identify lumbar disorders including Herniated Disk, Spinal Stenosis, Spondylosis, and Healthy instances by applying several architectures, including CNN, MobileNet, DenseNet, VGG16, AlexNet, ResNet, and a unique hybrid CNN-SVM model. By attaining 100% classification accuracy with excellent precision, recall, and F1-scores across all categories, the hybrid model in particular beat all other models. Standard classification measures, confusion matrices, training-validation accuracy/loss curves, and Grad-CAM heatmaps for interpretability were used to evaluate each model ‘s performance. DenseNet and other models suffered from overfitting and inadequate minority class detection, but older CNN models demonstrated great generalization. MobileNet and other light weight designs were efficient, but they had trouble with class imbalance. On the other hand, the hybrid CNN-SVM model produced better classification capabilities by employing SVM to improve decision boundaries and capture pertinent spatial variables. In addition to demonstrating the promise of ensemble and hybrid deep learning approaches in the field of medical imaging, this study demonstrates how incorporating visualization methods such as Grad-CAM greatly enhances clinical interpretability, which is essential for practical implementation.

References

[1] J. Hai, J. Chen, K. Qiao, N. Liang, Z. Su, H. Lv, and B. Yan, “Semantic Contrast with Uncertainty-Aware Pseudo Label for Lumbar Semi-Supervised Classification,” 2024. [2] K. (Kathy) Chen, L. Zheng, H. Zhao, and Z. Wang, “Deep Learning-Based Intelligent Diagnosis of Lumbar Diseases with Multi-Angle View of Intervertebral Disc,” 2024. [3] B. Mulugeta Abuhayi, Y. A. Bezabb, and A. M. Ayalaew, “Lumbar Disease Classification Using an Involutional Neural Based VGG Network,” 2023. [4] B. Atalay, Y. U. Yolcu, S. N. Magge, R. Whitmore, and Z. Ghogawala, “P261. High-grade lumbar interbody fusion correlates with improved patient related outcomes: a single center prospective registry study,” The Spine Journal, vol. 24, no. 9, pp. S192–S193, Sep. 2024, doi: 10.1016/J.SPINEE.2024.06.384. [5] G. Abdelmalek et al., “P44. Interobserver agreeability and clinical validity of the Miskin Classification System: a standardized radiographic assessment of lumbar degenerative disease,” The Spine Journal, vol. 24, no. 9, pp. S84–S85, Sep. 2024, doi: 10.1016/J.SPINEE.2024.06.065. [6] N. Löbig, C. Pagenkopf, M. Südmeyer, and K. Werheid, “Reduced awareness of cognitive deficits in idiopathic normal pressure hydrocephalus and its change following lumbar puncture,” Cortex, vol. 179, pp. 157–167, Oct. 2024, doi: 10.1016/J.CORTEX.2024.07.010. [7] T. Hirono et al., “Medication-based profiling of older orthopedic patients: a multicenter cross-sectional study,” BMC Geriatr, vol. 24, no. 1, pp. 1–10, Dec. 2024, doi: 10.1186/S12877-024-05284-8/TABLES/6. [8] H. Hu, C. Wu, and L. Tan, “A new classification of lumbar Basstrup’s disease based on radiographic and clinical symptoms,” European Spine Journal, vol. 33, no. 4, pp. 1347–1359, Apr. 2024, doi: 10.1007/S00586-024-08128-6/METRICS. [9] B. Kelly and E. Schulze, “A - 134 Consistent Failed Performance on Performance Validity Tests in Early-Onset Alzheimer’s Disease Confirmed by Lumbar Puncture,” Archives of Clinical Neuropsychology, Sep. 2024, doi: 10.1093/ARCLIN/ACAE067.148. [10] J. DiDomenico et al., “Does comorbid depression and anxiety portend poor long-term outcomes following surgery for lumbar spondylolisthesis? Five-year analysis of the Quality Outcomes Database: Presented at the 2024 AANS/CNS Joint Section on Disorders of the Spine and Peripheral Nerves,” J Neurosurg Spine, vol. 1, no. aop, pp. 1–10, Sep. 2024, doi: 10.3171/2024.5.SPINE24325. [11] T. Bassani, A. Colombini, L. Pallotta, L. M. Sconfienza, D. Albano, and M. Brayda-Bruno, “Association between MRI measurements of lumbar spine alterations and self-reported outcomes of pain and disability in subjects with non-specific low back pain,” European Spine Journal, pp. 1–9, Aug. 2024, doi: 10.1007/S00586-024-08449-6/METRICS. [12] R. Zhang, “A State-of-the-Art Survey of Deep Learning for Lumbar Spine Image Analysis: X-Ray, CT, and MRI,” AI Medicine, pp. 3–3, Jul. 2024, doi: 10.53941/AIM.2024.100003. [13] Z. Zhang and H. Shao, “Lumbar disc herniation typing method based on double-branch fusion and multiscale attention,” https://doi.org/10.1117/12.3034774, vol. 13210, pp. 863–868, Jul. 2024, doi: 10.1117/12.3034774. [14] K. Chen, L. Zheng, H. Zhao, and Z. Wang, “Deep Learning-Based Intelligent Diagnosis of Lumbar Diseases with Multi-Angle View of Intervertebral Disc,” Mathematics 2024, Vol. 12, Page 2062, vol. 12, no. 13, p. 2062, Jul. 2024, doi: 10.3390/MATH12132062.

Copyright

Copyright © 2025 K Gayathri, Karnam Bhagya Sree, Kathi Nikitha, K Supriya, Mr. N Vijaya Kumar, Dr. R Karunia Krishnapriya, Mr. V. Shaik Mohammad Shahil, Mr. Pandreti Praveen. 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.