Self-supervised learning assisted diagnosis for mitral regurgitation severity classification based on color Doppler echocardiography

Figures

• 

  Figure 1. Flow chart for data selection. Patient selection for testing, training, and validation was described. MR, mitral regurgitation.

• 

  Figure 2. Automatic self-supervised feature extraction framework. (A) Proxy task for color Doppler self-supervised feature extraction. (a) represents image transformation in data pre-processing, which includes random color distortion and Gaussian blur. (b) Patch rearrangement, which serves for structure recovery. Z1 and Z2 follows an equal rearrangement. (c) Siamese-Octad 2D ResNet-34 is employed for feature extraction from each single image patch, which leads to feature vector as output. (d) FC layer represents fully-connected layer, and it outputs the category possibility of each possible permutation. Lsr and Lc indicate structure recovery loss and color transform consistency loss, respectively. (B) Transfer to downstream multi-task network. (a) ROI cropping represents central area cropping. (b) Siamese-Octad 2D ResNet-34 is employed for feature extraction from each single video frame, which leads to feature vector as output. (c) LSTM captures the information of previous frames for better feature representation. (d) Skip connection represents feature concatenation of each corresponding outputs of green block and red block. (e) 2D segmentation decoder aims to decode low-level feature to…

• 

  Figure 3. Indicator illustration example. Six indexes (MR jet length/LA length, MR jet length, LA width, LA area, MR jet area, MR jet area/LA area) are evaluated by our self-supervised model. Green line represents MR jet area, dark blue line represents left atrial area, light blue lines represent LA width, MR jet length, LA length. MR, mitral regurgitation; LA, left atrium.

• 

  Figure 4. Box plot figure for six indexes. The relatedness between each index and ground truth. MR, mitral regurgitation; LA, left atrium.

• 

  Figure 5. Performances of six indexes generated by the AI segmentation model in classification of moderate-severe vs. non-moderate-severe in MR patients based on ROC curves. AI, artificial intelligence; MR, mitral regurgitation; ROC, receiver operating characteristic.