A stacking based deep learning framework integrating random search neural architecture search for meniscus tear diagnosis

BACKGROUND: Accurate and rapid diagnosis of meniscal tears is crucial for effective management of sports-related injuries and degenerative knee disorders. Magnetic resonance imaging (MRI) is widely used for meniscus evaluation; however, manual interpretation is time-consuming and subject to inter-observer variability. Automated and reliable classification systems may therefore support clinical decision-making. PURPOSE: This study aims to develop a robust and interpretable deep learning framework for the automatic four-class classification of meniscal conditions using MRI images. METHODS: A total of 2,000 knee MRI images obtained from a tertiary care university hospital were categorized into four classes: Grade I, Grade II, Grade III, and normal meniscus. A Random Search-based Neural Architecture Search (RS-NAS) strategy was used to generate task-specific CNN architectures. The class probability outputs of the top-performing NAS-derived CNN models were combined using a late-fusion stacking strategy, with ElasticNet employed as the meta-learner. The proposed framework was evaluated using both a single-run assessment and a repeated validation protocol consisting of 10 independent runs with 5-fold cross-validation. Performance was assessed using ACC, F1-score, MCC, AUC, and PR-AUC. Statistical comparisons and explainability analyses, including Grad-CAM and ElasticNet feature importance, were also performed. RESULTS: In the single-run evaluation, the ElasticNet-based stacking model achieved an ACC of 0.9300, F1-score of 0.9304, and AUC of 0.9913. Under the repeated 10-run 5-fold cross-validation protocol, the proposed model obtained an ACC of 0.9083 ± 0.0019, F1-score of 0.9084 ± 0.0018, MCC of 0.8783 ± 0.0024, AUC of 0.9883 ± 0.0008, and PR-AUC of 0.9719 ± 0.0020. The proposed stacking framework outperformed the best NAS CNN, DenseNet121-TL, averaging ensemble, and majority voting ensemble across the main evaluation metrics. Grad-CAM visualizations and feature importance analysis further indicated that the model relied on clinically meaningful image regions and informative base-model predictions. CONCLUSIONS: The integration of RS-NAS with ElasticNet-based late-fusion stacking provides a stable, high-performing, and interpretable framework for four-class meniscus MRI classification. Although the single-run evaluation showed higher peak performance, the repeated cross-validation results provide a more reliable estimate of model robustness and generalization. Future studies should validate the proposed framework using independent multi-center datasets and different imaging protocols.