Preprocessing

A radiologist with over 5 years of experience manually segmented the hypoperfused regions on the CTPA images for the 25 patients in group A, using V/Q scan images as a reference for segmentation. To appreciate the minor differences in the parenchymal attenuation, grey-level mapping with a window width ranging from 200 to 900 and a window level ranging from -1,100 to -800 was used. Only the regions of hypoperfusion visible in the CTPA were segmented. Regions where hypoattenuation was suspectedly caused by lung disease other than CPE were not segmented. Areas with relevant artefacts relating to, e.g., motion or beam hardening, were left unsegmented. Regions smaller than 5 mL were left unsegmented. Hyperattenuating lesions relating to the CPE were included in the segmentation, e.g., lung infarctions and atelectasis in the diseased area if seen in the hypoperfused region. Segmentation was done on the axial images using the image processing and visualisation platform 3D Slicer [30].

Due to a large number of axial images, with a median of 335 (min–max, 111–570) per CTPA, we used a tool based on morphological contour interpolation for assistance in segmentation [31], which allowed any number of slices to be skipped between slices segmented manually. This tool interpolates the manually segmented slices by detecting and aligning overlapping shapes in the adjacent segmented slices and generating a transition sequence of one-pixel dilatation between the overlapping shapes, and a median of this sequence is taken as a result to fill the skipped slices [31, 32]. The radiologist segmented the hypoperfused regions to every 3–10 axial images by hand, and the complete segmentation was created by the interpolation tool. Contours of the interpolation were evaluated and approved by the radiologist before committing to the segmentation. If the radiologist disagreed with the interpolation, the contours of the interpolated segmentation were modified and corrected manually. In group A, segmentation was done in this manner to every axial CTPA image for the CNN training and analysis. In group B, there was no manual segmentation done for hypoperfusion. Additionally, in both groups, the lungs were segmented as a whole, and the CTPA image data outside of the lungs was excluded from the analysis.

We applied a 48%–12%–40% training-validation-test split, i.e., 12 CTEPH positive and 12 negative CTPA volumes for training, 3 positives and 3 negatives for validation, 10 positives, and 10 negatives for testing. We split the data temporally: all the training and validation data had been imaged before all the test data. The hospital ethics committee approved the study. Informed consent was waived because of the retrospective design and anonymous clinical data used for the analysis.