1349003 - Deep learning for automated insertion point annotation of CMR late gadolinium enhancement and virtual native enhancement images

Background: Late gadolinium enhancement (LGE) on cardiovascular magnetic resonance (CMR) imaging can reproducibly identify the location and the extent of myocardial scar¹. Left ventricular (LV) segmental analysis that allows accurate assignment of myocardial territories (based on the AHA 16-segment model²) is particularly important for assessment of coronary artery disease and aiding clinical decision-making regarding revascularisation. Advances in automated LGE post-processing have primarily focused on segmentation of the LV endo- and epi-cardial contours³; however annotation of the anterior right ventricular (RV) insertion point to determine the orientation and segment assignments is also critical to allow full segmental quantification of scar volume fraction and transmurality on LGE. We present a deep learning approach for automated anterior RV insertion point annotation for LGE, applicable to the emerging contrast-agent-free CMR modality, virtual native enhancement (VNE) imaging⁴.

Methods: We adapted our CMR tracking technology⁵ to determine the RV entry line, as defined by the anterior RV insertion point and LV centre point. In short, the method involved a residual neural network backbone to regress the Cartesian coordinates of both points in a given image⁵. The imaging data comprised of 1085 (n = 816/269 LGE/VNE) short-axis images from 135 patients with myocardial infarction. The datasets were manually annotated by 2 experienced observers, which served as ground truth. A randomly selected set of 140 images was used for independent testing, and the rest for training the deep neural networks. The spatial annotation accuracy was measured by: (i) the Euclidean distance (mm) between the reference and predicted points, and (ii) the absolute angular error (º) between the ground-truth and predicted RV entry lines, in all slices and stratified by the slice position. The segmental scar transmurality error was measured by comparing the metrics derived from the ground-truth and predicted annotations on available LV myocardial manual contours (n = 70), using the intra-class correlation coefficient (ICC).

Results: The proposed model accurately predicted the RV entry line in all of the testing materials (Fig. 1), taking around 60 ms per slice. The overall Euclidean distance error was 2.9 ± 2.0 mm, the angular distance error 6.2 ± 5.9º, with an increasing trend towards the apical slices (Table 1). The errors between the test LGE set (n = 105; 2.8 ± 2.0 mm, 5.9 ± 5.7º) and the test VNE set (n = 35; 3.2 ± 2.0 mm, 6.9 ± 6.4º) were not statistically significant (p > 0.05). There was no perceivable impact on the derivation of the segmental scar transmurality (Fig. 2; n = 70, bias = 0.14 ± 3.49%, ICC = 0.99).

Conclusion: The presented deep learning-based approach is robust and effective for annotating the anterior RV insertion point in CMR images. Together with LV segmentation and quality control, it completes the pipeline for automated segmental quantification of LGE and VNE for clinical use.