Automated lumbar intervertebral disc identification and herniation detection in MR images using cascade CNN architecture

Abstract

Objective

Identifying herniated discs in MRI scans is inherently challenging due to the small size, irregular shape, and complex appearance of the affected regions. Conventional approaches typically rely on semi-automated region-of-interest (ROI) selection and single-model classification using either axial or sagittal views, limiting diagnostic performance. This study aims to develop an automated, accurate, and efficient system for the detection and classification of lumbar intervertebral disc herniation using deep learning models applied to axial and sagittal MR images.

Methods

A YOLO-based framework was developed to automatically identify lumbar intervertebral discs (IVD_1-5) and extract ROIs from MR images. Attention-enhanced and fine-tuned VGG19 and ResNet50 models were employed to analyze axial and sagittal images for herniation detection. A decision fusion strategy was used to combine the classification probabilities from both models to further enhance accuracy. The dataset underwent extensive preprocessing and augmentation to improve model robustness and generalization.

Results

The proposed approach demonstrated exceptional performance in detection and classification tasks. For detection, the model achieved mAP50 scores of 95.18 % (axial IVD_1-5), 99.50 % (lumbar regions), and 94.87 % (sagittal IVD_1-5). Classification accuracy reached 97.05 % for axial images and 97.45 % for sagittal images, increasing to 98.09 % with decision fusion.

Conclusion

Designed to assist physicians, especially during high-demand periods such as pandemics, this approach has the potential to improve diagnostic efficiency and reduce clinical workload.