Abstract: BACKGROUND: Knee finite element modelling can provide insight into knee mechanics, but its complex image segmentation is more difficult for researchers. With the development of deep learning techniques, deep learning techniques have been widely used in knee joint finite element modelling. 
OBJECTIVE: To replace the manual segmentation step in finite element modelling of the knee joint by using 3D Swin UNETR in combination with a semi-automatic segmentation technique for statistical shape models.
METHODS: Manual (artificial) knee joint finite element model was developed based on MR and semi-automatic knee joint finite element model was developed based on 3D Swin UNETR+ statistical shape model segmentation. The same loads and boundary conditions were applied to both models. Validation was performed by calculating the Dice similarity coefficient, mean distance, and comparing the peak equivalent stresses, maximum principal stresses, and maximum shear stresses of the two models.
RESULTS AND CONCLUSION: (1) The Dice similarity coefficients of the manual and semi-automatic segmented femur and tibia were more than 98%, and the average distances were less than or equal to (0.35±0.08) mm. (2) With the longitudinal load of 750 N and 10 Nm internal overturning moment applied to the femur tip of both manual and semi-automatic finite element models, the peak equivalent stress, maximum principal stress, and maximum shear stresses of meniscus in manual finite element model were 14.12, 18.54, and 7.35 MPa; peak equivalent force, maximum principal stress, and maximum shear stress of femoral cartilage were 2.22, 2.15, and 1.18 MPa; peak equivalent force, maximum principal stress, and maximum shear stress of tibial cartilage were 2.50, 1.91, and 1.41 MPa; semi-automatic finite element model of meniscus: peak equivalent force, maximum principal stress, and maximum shear stress were 14.93, 18.53, and 7.75 MPa. The peak equivalent force, maximum principal stress, and maximum shear stress of femoral cartilage were 2.26, 2.18, and 1.20 MPa; the peak equivalent stress, maximum principal stress, and maximum shear stress of tibial cartilage were 2.60, 1.91, and 1.46 MPa. The peak equivalent stress, maximum principal stress, and maximum shear stress of manual and semi-automatic finite element models were basically consistent, with no significant difference (P > 0.05). (3) The semi-automatic segmentation technique proposed in this study can replace manual segmentation in creating accurate finite element models of the knee joint.

Key words: knee joint, finite element model, 3D Swin UNETR, statistical shape model, semi-automatic segmentation, manual segmentation