Meniscal tears and degenerative changes are the most common pathologies affecting the knee joint. In magnetic resonance imaging (MRI), these lesions often manifest at small spatial scales with indistinct boundaries and heterogeneous intensity patterns, making accurate automated analysis challenging. To address these difficulties, this paper proposes CTransFuse, a compact and efficient hybrid CNN-Transformer framework designed for lesion-aware meniscus segmentation and lesion identification. The proposed model adopts a parallel dual-branch design that leverages the strengths of convolutional neural networks in the ability of Transformers to model global contextual reasoning. Specifically, a lightweight multi-scale pyramid Transformer encoder (CosFormer) is introduced to efficiently extract global representations, while a bidirectional attention fusion module (BiFusion) is introduced to enables effective cross-level feature interaction and progressive refinement across different semantic levels. This design improves discriminative feature learning while maintaining low computational cost. The experimental results demonstrate that CTransFuse operates at an inference speed of 177.6 frames per second with only 12.6M parameters and 20.8G MACs. The model achieves competitive segmentation performance, yielding Dice scores of 79.28% and 82.57%, and mean intersection-over-union (mIoU) scores of 83.25% and 87.18% for meniscal tears and degenerative changes, respectively. For lesion classification, CTransFuse achieves an area under the ROC curve (AUC) of 97.45% and an area under the precision-recall curve (AUPR) of 97.49%, demonstrating robust under class-imbalanced conditions. Additional experiments on the public MRNet dataset demonstrate the framework's robustness and strong generalization capability of the framework. These results highlight the potential of CTransFuse for intelligent knee MRI analysis and computer-aided clinical assessment.
Li et al. (2026) studied this question.