Fig.1

The overall block diagram of the proposed cascaded network. Let the coordinate of the lower left corner of the original image be (0, 0), (199, 204) is the center coordinate of the extracted cartilage and meniscus, (104, 109) and (295, 300) are the coordinates of the lower left and upper right corners of the crop area in the original image, respectively. The final result is obtained by merging the segmentation results of the two networks according to the extracted position information, and post-processing

Fig.2

The structure of the first U-Net in the proposed cascade network

Fig.3

The first row is the initial image and the corresponding segmentation results of cortical bone, cancellous bone, muscle, fat, and background. Extract the cavity in the background and calculate its centroid to obtain the position information of cartilage and meniscus. Cut out with the centroid as the center to obtain a sub-image containing cartilage and meniscus, which is the input image of the second network. The second row respectively shows the cut-out schematic diagram and the segmentation results of cartilage and meniscus with the second network

Table 1

Analysis for the segmentation evaluation indicators of the proposed method and 4 comparison methods

Fig.4

The comparison between the final segmentation results obtained by using the 5 automatic segmentation methods and the manual segmentation. From top to bottom, 4 consecutive equally spaced slices are collected for a volunteer image. Among them, fat, muscle, cancellous bone, cortical bone, cartilage, and meniscus are marked as yellow, red, white, blue, green, and pink, respectively

Table 2

Comparison of the relative errors of maximum SAR10g between the 5 automatic segmentation methods and the manual segmentation

Fig.5

SAR10g distribution maps of a volunteer's knee joint. From left to right, they are based on the manual segmentation, and the models obtained by the proposed method, the U-Net using cross-entropy loss, the U-Net using focus loss, the U-Net using attention mechanism and U-Net++, respectively. The SAR10g values of the six models are 1.309 8 W/kg, 1.304 5 W/kg, 1.402 4 W/kg, 1.316 2 W/kg, 1.383 5 W/kg, 1.506 0 W/kg. The first row to the sixth row are the SAR10g distribution maps of 6 uniformly distributed slices, and the last row is the maximum density projection of the SAR10g