Comparison of manual and automatic measurements
The algorithm performance was indistinguishable from human observers’ performance. The automatic algorithm successfully segmented the LVOT and aortic root and measured the aortic annulus diameter in all 16 images, with mean computation time 9.9 s. Interobserver correlation coefficient for the manual measurements was comparable to that reported by others [
4]. The correlation and deviation between the automatic and each manual measurements were comparable to the interobserver reliability.
Perimeter-derived measurements showed the closest agreement with the manual observers. Since the perimeter-derived diameter changes the least during the cardiac cycle [
8], this measurement should not be affected by errors in mid-systole detection. This is therefore a more robust measurement compared to area, major and minor diameters, and was chosen as the algorithm’s main output diameter.
The annulus plane is normally defined as the plane spanned by the hinge points of the three valve cusps. However, the hinge points are not explicitly defined in the described model. Because the disk representing the aortic valve is non-deformable and shares the same pose transform as the surface model, the detected annulus will align perpendicular to the long axis of the LVOT.
Since the hinge point plane and the perpendicular plane are closely aligned, we propose that a perpendicular plane is a good estimation of the anatomical annulus plane. In the rare cases where these planes are not aligned, we submit that a perpendicular plane is of clinical relevance since a prosthetic valve is more likely to align with the LVOT long axis than the native valve’s hinge points.
The largest absolute deviation between the two manual observers was 5.4 mm. Poor image quality, low frame rate (7.7 vps) and a wide sinus of Valsalva lead to the significant interobserver deviation. The largest absolute deviation between automatic and manual measurements was 5.5 mm. In this case, the automatic method grossly overestimated the annulus diameter, resulting from a very wide sinus of Valsalva combined with dropouts close to the annulus. Although these discrepancies would indicate a difference in device selection, it is extremely unlikely that these specific images would be used as the basis of device selection.
Two recordings had visible stitching artifacts. In both recordings all absolute deviations between the the automatic and manual measurements were ≤0.6 mm, indicating robustness against stitching artifacts.
Recently, a validation study of the first description of an automated aortic root modeling and quantification algorithm for 3D TEE images was published [
9]. The study reported annulus diameter agreement (mean ±SD) of 1.1±1.3 and 3.6±2.3 mm for sagittal and coronal diameters respectively, which is comparable to our results. However, manual identification of peak systole and end diastole was required, and manual segmentation adjustments were needed in 23 of 69 TEE recordings. The reported interobserver variability was 0.2±0.56 and 0.0±0.61 mm. Although our presented algorithm is fully automatic and therefore has no interobserver variability, these values are comparable to our sensitivity analysis results. The reported computation and adjustment time was 2.3±0.6 minutes, which is significantly longer than our results.
This method is based on machine learning and statistical shape models [
10]. However, these algorithms require a large database of recordings annotated with manual ground truth segmentations. The presented method is simpler and does not rely on a history of previous segmentations.
Within the presented framework, there are several quality measures available that can potentially be used to automatically identify poor segmentations, e.g. number of discarded edge profiles, deviation from segmentation surface to detected edges or the state covariance estimates. These should be further investigated with an available gold standard to create criteria that can automatically judge the segmentation quality.
This study used a limited sample size of 16 patients. Further studies with a larger number of patients should be performed.
Comparison of measurements in prospective 3D TEE images with ECG gated Multi-Slice Computed Tomography (MSCT) gold standard should be carried out to investigate if the algorithm can render MSCT superfluous for a significant portion of TAVI candidates.