Comparative analysis of model complexity.

Knee Ailments, such as meniscus injuries, bother millions globally, with research showing that more than 14% of the population above 40 years lives with meniscus-related conditions. Conventional diagnosis techniques, like manual MRI interpretation, are labour-intensive, error-prone, and dependent on skilled radiologists, making an automatic and more accurate alternative indispensable. Current deep-learning solutions heavily depend on CNNs, which perform poorly in long-range dependencies and global contextual info. We proposed MV2SwimNet, a hybrid of MobileNetV2 and Swin Transformer, integrating Window Multi-Head Self-Attention (W-MSA) and Multi-Stage Hierarchical Representation (MSHR), efficiently incorporating both local and global features towards enhanced diagnostic capability. Our strategy utilizes the efficiency of lightweight MobileNetV2 coupled with a hierarchical architecture and self-attention-based Swin Transformer, enabling better spatial representation and advanced feature extraction. W-MSA allows our model to process MRI scans effectively by attending to the corresponding regions of images. In contrast, MSHR adjusts feature representations across different levels in a way that allows for progressive and robust learning in stages. We tested MV2SwimNet on two sets using 3-fold cross-validation and achieved 99.94% and 96.04% accuracy on dataset1 and dataset2, which beats state-of-the-art techniques. These results confirm MV2SwimNet efficiency, robustness, and real-world application potential in medicine, providing a highly accurate, automated medical diagnosis tool for knee disease detection. The code of the proposed method can be accessed through the URL: https://github.com/Visheshtanwar/MV2SwimNet