Automated Recognition of Ultrasound Cardiac Views Based on Deep Learning with Graph Constraint

Figures

• 

  Figure 1. An illustration of the proposed classification framework. (a) The sample with similar samples defined by the similar graph are inputted into the same CNN, and the representation produced by similar samples is used as graph regularizer to compute unsupervised loss, which are combined with supervised loss to train the CNN. (b) The CNN includes three modules. STM learns the six affine parameters of the cardiac deformations to reduce the geometric distortion. The Inception-V3 with output layer removed is used for feature extraction. The extracted feature maps are inputted to the SE, which adaptively recalibrates channel-wise feature responses, and then the predicted probability is outputted.

• 

  Figure 2. The flowchart of spatial transformer network. STM receives original images and uses a localization network to learn the parameters of the geometric transformation during the cardiac cycle. This network consists of two layers. First, the convolution layer of 5 × 5 filters is used to output 128 feature maps, followed by a global averaging layer and a fully connected layer. The localization network produces 6 parameters of affine transformation. The transformation parameters are in turn applied to every inputted image to perform a geometric transformation and reduce cardiac deformations.

• 

  Figure 3. The flowchart of SE. As a feature extractor, Inception-V3 is removed from the final output layer, and the outputted feature maps with 2048 channels are fed into SE, where the channel attention mechanism is introduced to enhance the channel that are more effective for accurate predictions.

• 

  Figure 4. Confusion matrices. (a) Testing set in Dataset 1, (b) Testing set in Dataset 2.