Fig. 3

CMG Net is effective for segmentation of normal hip joints. A Comparison of learning curves from the training data with the proposed CMG Net and other alternative Net (Since we tune the class weights for different networks to ensure its performance so they don’t have to converge to one loss during training process.) B Qualitative comparison of the segmentation results obtained by the automatic segmentation to manual segmentation on a given axial slice of the normal hip joint. C Quantitative comparison of the segmentation results obtained by the automatic segmentation to manual segmentation on the normal hip joint. D After we piled up all the segmented layers according to the original CT sequence of a normal hip, we could rebuild an accurately segmented 3D hip model. A 3D model rebuilt by original CT images; B 3D model rebuilt by CT images segmented by CMG Net; C the anatomical sturctures of both femur and acetabular can be observed clearly. ***p < 0.001, ****p < 0.0001 versus CMG Net