Regressive vision transformer for dog cardiomegaly assessment

Figures

• 

  Figure 1. Our calculation of VHS, there are six key points (A, B, C, D, E and F). VHS = 6 × (AB + CD)/EF.

• 

  Figure 2. The overall architecture of our proposed regressive vision transformer (RVT) model. We first employ a pyramid vision transformer as an encoder to extract low and high-level features. Then, we add a feature fusion module to predict six key points of the VHS score. An orthogonal layer is proposed to ensure the perpendicularity between line segments AB and CD as shown in Fig. 1.

• 

  Algorithm 1. Regressive Vision Transformer (RVT). \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$B(\cdot )$$\end{document}B(·) denotes the mini-batch training sets, and I is the number of iterations.

• 

  Figure 3. Data distribution of training, validation, test, and all DogHeart dataset, respectively.

• 

  Figure 4. Six sample dog X-ray images of our collected DogHeart dataset.

• 

  Figure 5. The training convergence of 13 different models.

• 

  Figure 6. Predicted results of our RVT model. The red color lines are the ground truth lines and the blue color lines are the predicted results.

• 

  Figure 7. Predicted results comparison. The red color lines are the ground truth lines and the blue color lines are the predicted results.

• 

  Figure 8. Confusion matrices of validation and test datasets. The X-axis is ground truth labels, and the Y-axis is predicted labels. The numbers (e.g., 30, 2, 0) are the predicted class numbers. Black percentages \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$e.g., 30/200 = 15\%$$\end{document}e.g.,30/200=15% are the ratio between the number of predicted labels and the total number of images. The green percentages are the correct predicted percentages, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$e.g., 30/33 = 90.9\%$$\end{document}e.g.,30/33=90.9%, and the red percentages are the wrong predicted percentages, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt}…

• 

  Figure 9. The boxplot of improvement after using the orthogonal layer.