基于基础模型编码器多尺度因果干预的CT功能性骨髓分割OA

Automatic segmentation of functional bone marrow in three-dimensional computed tomography(CT)images is often challenged by low grayscale contrast,blurred boundaries,discrete spatial distribution,and pronounced inter-individual variability.To address these issues,a three-dimensional CT functional bone marrow segmentation model based on multiscale causal intervention with a foundation model(UF-MedSAM3D)is proposed.The model adopts a U-shaped encoder-decoder architecture as the backbone and incorporates a frozen MedSAM3D foundation model encoder to extract multiscale global semantic features.Furthermore,multiscale causal intervention modules are designed within the encoding stage to enable causal regulation and efficient fusion of global semantic information and local structural features,thereby suppressing spurious correlations and enhancing boundary representation capability.The proposed method is evaluated on a self-constructed functional bone marrow database(FBMD).The Dice coefficient,mIoU,HD95,and ASSD for the FBM1 region reach 81.02%,72.08%,5.347,and 0.941,respectively,while those for the FBM2 region reach 83.11%,74.11%,4.833,and 0.542,respectively.Experimental results demonstrate that,compared with mainstream three-dimensional segmentation models such as UNETR,SwinUNETR,TransBTS,nnFormer,and nnUNetv2,the proposed method achieves superior segmentation accuracy and boundary consistency.