Fig. 1

The whole flowchart of knee joint model construction using the proposed method

Fig. 2

The architecture of the proposed CGAN. The countermeasure loss is calculated according to the 0/1 matrix of the output of the discriminator network. The error is propagated back to the generator network and the discriminator network, and the L1 loss is calculated by using the artificially labeled image and the generating result of the generator network, transmitting back to the generator network, and the network parameters are continuously adjusted according to the gradient descent principle

Fig. 3

The structure of generator network. The network input is the original gray-scale image, and the output consists of muscle, fat, bone and background images

Fig. 4

The attention modules used in generator network. The attention coefficient matrix α is calculated by using the decoded partial characteristic graph dl and the coded partial characteristic graph el-1 as the input, el-1 multiplies the attention coefficient matrix α pixel by pixel to select the focus region

Fig. 5

The structure of discriminator network

Table 1

Evaluation indexes of the segmentation results of various methods (the proposed method, U-Net, Attention U-Net compared with the manual labeling results)

Fig. 6

Segmentation results of one knee section of a female volunteer based on different methods, where yellow represents fat, red represents muscle, and white represents bone, with a field of view of 150 mm×150 mm

Fig. 7

Models of the knee joint of a female volunteer constructed with different methods. (a) artificial labeling; (b) the proposed method；(c) method 1; (d) method 2; (e) method 3

Table 2

Statistics of the relative error of the maximum SAR10g (the proposed method and three contrast methods compared with the results of artificial labeling)

Table 3

Statistics of the maximum SAR10g (the Virtual human)

Fig. 8

SAR10g distribution of the same volunteer’s knee. From top to bottom are the distribution of five consecutive MR slices, from left to right, are the results based on artificially labeling, the proposed method, method 1, method 2 and method 3 respectively

Fig. 9

The same volunteer’s maximum intensity projection of SAR10g by (a) artificially labeling; (b) the proposed method; (c) method 1; (d) method 2; (e) method 3