Lightweight Multi-View Probabilistic Fusion for Knee MRI Classification

Abstract

Accurate diagnosis of knee pathologies from Magnetic Resonance Imaging (MRI) remains a challenging task, particularly in the presence of class imbalance and multi-planar anatomical variability. Although deep learning approaches have achieved remarkable success, they often require substantial computational resources and large annotated datasets, limiting their deployment in resource-constrained clinical environments.

In this study, we propose a lightweight multi-view machine learning framework for automated knee MRI diagnosis. Instead of end-to-end deep learning training, we employ pretrained convolutional neural networks for feature extraction and combine them with classical machine learning classifiers. Sagittal, axial, and coronal MRI views are processed independently to capture complementary anatomical information. Their probabilistic outputs are integrated using a weighted ensemble fusion strategy. To address the trade-off between precision and recall in imbalanced datasets, an adaptive threshold optimization approach based on the precision–recall curve is introduced to maximize the F1-score without degrading Area Under the Curve (AUC) performance.

Experimental evaluation on multi-view knee MRI data demonstrates that the proposed framework achieves AUC scores of 0.924 for abnormality detection, 0.902 for ACL tear detection, and 0.749 for meniscus tear detection. Results indicate competitive diagnostic performance while maintaining significantly lower computational complexity compared to fully end-to-end deep learning models.