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

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Diagnostic Support
• Radiology and Orthopedics

Background:

• Accurate diagnosis of meniscal tears is critical for managing knee injuries and disorders.
• Magnetic resonance imaging (MRI) is standard for meniscus evaluation but manual interpretation is time-consuming and variable.
• Automated classification systems are needed to support clinical decision-making.

Purpose of the Study:

• To develop a robust and interpretable deep learning framework for automatic four-class classification of meniscal conditions using MRI images.
• To enhance diagnostic accuracy and efficiency in meniscus tear detection.

Main Methods:

• A dataset of 2,000 knee MRI images was classified into four categories: Grade I, Grade II, Grade III, and normal meniscus.
• A Random Search-based Neural Architecture Search (RS-NAS) strategy generated task-specific Convolutional Neural Network (CNN) architectures.
• Late-fusion stacking combined top CNN models using ElasticNet as a meta-learner, evaluated via repeated 10-run 5-fold cross-validation.

Main Results:

• The ElasticNet-based stacking model achieved high accuracy (ACC: 0.9083 ± 0.0019) and Area Under the Curve (AUC: 0.9883 ± 0.0008) under cross-validation.
• The proposed framework outperformed individual NAS CNNs, averaging ensembles, and majority voting ensembles.
• Explainability analyses (Grad-CAM, ElasticNet feature importance) confirmed the model's reliance on clinically relevant image features.

Conclusions:

• The integrated RS-NAS and ElasticNet stacking framework offers a stable, high-performing, and interpretable solution for meniscus MRI classification.
• Repeated cross-validation demonstrated the model's robustness and generalization capabilities.
• Future validation on multi-center datasets and diverse imaging protocols is recommended.