Skip to main content
Fig. 3 | BMC Musculoskeletal Disorders

Fig. 3

From: Automatic segmentation model of intercondylar fossa based on deep learning: a novel and effective assessment method for the notch volume

Fig. 3

Illustration of the segmentation model. The U-Net network architecture is structured into an encoder and a decoder. The encoder follows the classic architecture of the convolutional neural network, with each convolutional blocks followed by a rectified linear unit (ReLU) and a maximum polling operation to encode image features at different levels of the network. The decoder up-samples the feature map with subsequent up-convolutions and concatenations with the corresponding encoder blocks. The network has two inputs: the preprocessed original MR images and the outlined ROI of the intercondylar fossa. The segmentation architecture consists of 10 convolutional layers, 11 residual blocks (RBs), two pyramid pooling modules (PSPPooling), five upsampling layers, six combine blocks and a fully connected (FC) layer, and a sigmoid layer. Conv = convolution, RB = residual block, PSPpooling = pyramid scene parsing pooling, FC = fully connected

Back to article page