Machine learning assisted reflectance spectral characterisation of coronary thrombi correlates with microvascular injury in patients with ST-segment elevation acute coronary syndrome

Figures

• 

  FIGURE 1. Experimental design.

• 

  FIGURE 2. Plot of the thrombus area within each spectral image (expressed as a fraction) against Thrombosis In Myocardial Infarction (TIMI) thrombus score. A threshold thrombus area fraction of 0.22 was used as the minimum value to define successful thrombectomy in patients with thrombus scores of 4 and 5.

• 

  FIGURE 3. Example hyperspectral images for two thrombus samples. Each image pixel contains a 150-wavelength reflectance spectrum which characterises the composition of the material being imaged. Example spectra for the marked pixels are shown on the right of the figure.

• 

  FIGURE 4. Example output of the k-means clustering analysis for two patients identified as having low and high index of coronary microvascular resistance (IMR) (centre panel) and small and large microvascular obstruction (MVO) (right panel), respectively. In the left panel, images (A,E) are photographs of the thrombus samples for each patient; images (B,F) show the results of the K = 11 k-means clustering analysis, with each colour corresponding to a different cluster; images (C,G) are the same as images (B,F) but with all non-thrombus pixels set to zero; and images (D,H) show the results of a second k-means clustering analysis with K = 7, performed only on thrombus pixels.

• 

  FIGURE 5. Example output of the basis function fitting process. The image on the left is a photograph of the sample. The false colour images show the fitted contributions to the spectral image from the filter, water ice, plasma, and red blood cells.

• 

  FIGURE 6. Linear regression analysis of correlations between thrombus spectral parameters predicted microvascular injury and actual microvascular injury indices for the thresholded data set of Oxford Acute Myocardial Infarction (OxAMI) samples. The plots show fits of the thrombus pixel k-fractions determined from the spectral images in the thresholded data set to (A) Eq. 2; (B) Eq. 4; and (C) Eq. 5. Note that the fitting coefficients ck are different for each fit. Sample number n and R2 value are shown for each correlation. Equivalent plots for the full OxAMI data set can be found in Supplementary Figure 3.