Abstract: To resolve the problems of limited robustness and poor adaptability to complex anatomical boundaries in knee CT images segmentation,an improved MedSAM-based automatic segmentation method is proposed.This method retains the MedSAM encoder to extract deep semantic features and designs an automated decoder that integrates hierarchical semantic aggregation with multi-scale skip connections,achieving global anatomical modeling and local boundary detail recovery.In addition,optimization strategies including slice position embedding,adjacent slice consistency constraints,and boundary awareness are integrated to enhance the model’s stability and generalization.Experimental results on a self-constructed dataset demonstrate the proposed method achieves Dice,IoU,Recall,and Precision scores of 0.984±0.004,0.970±0.005,0.986±0.004,and 0.982±0.005,respectively.The performance outperforms comparative models across all metrics,providing a reliable imaging foundation for precise preoperative planning and robot-assisted surgical navigation in ACLR.