A Detection-Driven Two-Stage Approach for Knee Cartilage MRI Image Segmentation with Memory Enhancement

Deep learning methods have made significant progress in medical image analysis, particularly in the segmentation of knee cartilage MRI images. However, the cartilage regions are of relatively small size and they are susceptible to background interference. As a result, traditional segmentation methods are prone to practical limitations such as class imbalance and background noise. The boundaries of knee cartilages are often unclear, especially in pathological conditions where the morphology of the cartilage may change, making it difficult to segment the boundaries of the cartilages accurately. To address these issues, this paper proposes an innovative two-stage approach that leverages the functions of object detection and segmentation (D2MSeg). At the first stage, an object detection module is used to precisely localize the knee cartilage region, thereby reducing background interference and the influence of non-target areas, effectively tackling the class imbalance problem. At the second stage, a memory-augmented module is proposed to help the model learn the fine details of the cartilage boundary by guiding the network to focus on fine-grained information. Experimental results on our dataset show that D2MSeg achieves a Dice score of 83.34%, outperforming most existing methods, including U-Net and TransUNet under comparable settings. These results highlight the effectiveness of our approach in capturing fine cartilage structures and its strong potential for clinical application.