Multi-feature Fusion based Deep Learning Framework for Enhanced Intervertebral Disc Degeneration Classification

Abstract

This study presents a robust framework for Intervertebral Disc (IVD) degeneration classification by incorporating deep Convolutional Neural Networks (CNNs) with handcrafted features. A hybrid descriptor set with morphological, edge-based, textural, and histogram based features was extracted to complement ResNet50-based deep features. The system was trained and evaluated on public lumbar spine MRI dataset, using a 5-fold stratified cross-validation scheme to ensure stable performance across degeneration stages. Comparative tests were conducted using Support Vector Machines (SVM), Random Forests, Gradient Boosting, and fully connected neural layers. The proposed CNN-handcrafted hybrid model achieved 98.3% accuracy, considerably outperforming the SVM baseline 80% and other alternatives. An ablation study demonstrated the critical and complementary role of each handcrafted feature group, with notable drops in performance when individual components were omitted. The fusion approach improves both interpretability and predictive strength, mainly in challenging cases like mild degeneration. This hybrid method offers a promising and explainable solution for clinicians, supporting additional reliable assessment of IVD degeneration from MRI data.