Artificial intelligence (AI) and computer vision have made great changes in medical image analysis in recent years. In musculoskeletal (MSK) medicine, MRI images provide detailed visualization of soft tissues, bones, and joints, which is important for accurate diagnosis and treatment planning. This research enhances the widely adopted nnU-Net framework by integrating Atrous Spatial Pyramid Pooling (ASPP) and Vision Transformers, resulting in two models: nnAtrousU-Net and nnAtrousTransformer. These improvements contribute to advancements in segmenting complex structures, such as thin and small cartilage, from MRI scans. Experiments were conducted on the OAI-ZIB dataset, including 507 knee MRI scans with manual annotations. Both models outperformed the baseline nnU-Net in key metrics such as Dice Similarity Coefficient (DSC) and average surface distance (ASD). Specifically, nnAtrousU-Net achieved a DSC of 90.46% for femoral cartilage and 86.91% for tibial cartilage, improving the state-of-the-art CNN-3D+SSM approach. The nnAtrousTransformer demonstrated competitive accuracy using half the parameters and one-third of the computational cost compared to nnAtrousU-Net. The proposed architectures offer improved segmentation accuracy, computational efficiency, and adaptability. It provides the solution for enhanced diagnostic and therapeutic applications in healthcare.