Mitral regurgitation detection and central/eccentric classification using transformer‐based deep learning in multi‐view echocardiography

Figures

• 

  FIGURE 1. Workflow of MR detection and centrality or eccentricity MR classification systems (same model framework).

• 

  FIGURE 2. ROC analysis of retrospective and prospective test sets in diagnosing MR. The closed lines represent the performance of the deep learning model on retrospective and prospective data sets respectively. Circles of different colors represent the performance of doctors with different seniority in diagnosing MR. The squares represent the levels of true positive rate and false positive rate when the regurgitation probability threshold is 0.5.

• 

  FIGURE 3. Activation maps of the attention‐based MR Detection model in Case 1 (A4C‐Color). (a) Self‐attention matrix covering 11 input frames, where index 0 corresponds to the [CLS] token. MR events occur between frames 1–6; row 0, highlighted by the red box, shows the attention weight assigned to each frame by the [CLS] token—warmer colors indicate higher attention (see color bar on the right). The model exhibits high attention to frames 1–6, consistent with clinically relevant MR periods. (b) Top row: Raw color Doppler echocardiographic frames from the MR sequence. Bottom row: Corresponding attention heatmaps superimposed on the same frames. The heatmaps show that the model primarily attends to the color Doppler information from the mitral valve orifice to the left atrium.

• 

  FIGURE 4. Activation maps of the attention‐based MR Detection model in Case 2 (PLAX‐Color). (a) Self‐attention matrix covering 11 input frames, where index 0 corresponds to the [CLS] token. MR events occur between frames 1–5; row 0, highlighted by the red box, shows the attention weight assigned to each frame by the [CLS] token—warmer colors indicate higher attention (see color bar on the right). The model shows high attention to frames 1–4, which happen to be MR frames. (b) Top row: Raw color Doppler echocardiographic frames from the MR sequence. Bottom row: Corresponding attention heatmaps superimposed on the same frames. The heatmaps show that the model can pay attention to the color Doppler information corresponding to the position of the mitral valve.

• 

  FIGURE 5. ROC analysis of retrospective and prospective test sets in diagnosing central‐eccentric MR.

• 

  FIGURE 6. Activation maps of the attention‐based central‐eccentric MR classification model (Case1‐central MR and Case2‐Eccentric MR). (a) and (c) Attention matrices across 11 input frames (frame 0 is [CLS] token) for two representative cases: The highlighted 0th row (top row, red box) indicates the attention weights assigned by the [CLS] token to each frame. Warmer colors (yellow—red) denote higher attention scores (see color bar on the right), while cooler tones (blue—green) indicate lower attention. In both cases, the model predominantly attends to the frame with maximal regurgitant jet intensity—consistent with clinically relevant timepoints. (b) and (d) Left: Original color Doppler echocardiographic images of the maximum regurgitation frame. Right: Corresponding attention heatmaps overlaid on the same frames show that the model can pay attention to the shape of MR in the left atrial surface.

• 

  FIGURE 7. Central Illustration (a) Using a single‐center retrospective database, we built two deep learning models based on Swin Transformer multi‐view diagnosis to detect MR and identify central regurgitation or eccentric regurgitation respectively; (b) Deep learning model workflow and internal validation in a retrospective test set. In addition, we compared the performance of the model with the diagnoses of physicians with different seniority; (c) Our model was also validated in a prospective test set.