Fig. 2.

An illustration of the two binary DA classifiers used in this study. (A) Two steps in the sinogram-based detection (SBD). First, one slice from a CT volume is thresholded and blurred, before being thresholded again to remove pixels in the body of the patient. The remaining pixels are thresholded again, revealing the streaks outside the patient’s body. The image is then transformed to the sinogram domain and the mean sinogram pixel intensity is computed. (B) An example of the ‘mean sinogram intensity’ for each slice in six CT volumes (each image represented with a different colour). A peak detection algorithm is applied to this plot for a given patient to detect slices likely to contain DAs. We annotate the detected slices with Xs to show that the algorithm detected one peak from each of the green and blue curves (both images labelled as ‘strong DA’). The dashed lines represent the peak detection threshold for each patient. (C) The CNN architecture used in the study. The network consisted of 5 convolutional layers (conv_1 to conv_5) creating a total of 64 filters. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)