Predicting Aortic Diameters with Deep Learning - Automated Annotation and Calculation of the Sinus of Valsalva and Sinotubular Junction
Length of the Aortic Sinus is a key consideration in identifying a number of aortic diseases. As such automating annotation of the aortic diameters can be helpful in its capacity to save time and, when paired in a more comprehensive clinical workflow, can be used to automatically identify patients that may be suffering from severe cardiac conditions. More importantly, however, relying on deep learning to produce aortic measurements can decrease inter-operator variability, especially in regards to the interpretation of the start and end of the endocardial border.
Vastly improving on previously published research, the iCardio.ai team achieves note-worthy results. Our team identified deficient aspects of the leading paper's approach to taking aortic diameters, with the primary improvement of our approach being a complete redesign of the underlying deep learning architecture.
The graph below projects the correlation of ground truth (GT) to predicted diameters calculated using an R2 coefficient. The diameter comparisons were based on pixel lengths, with images normalized to the size of the model input. The model can perform in both zoomed variations of the PLAX view, where the focus is on the aorta, as well as a more general view in the Standard PLAX perspective. Our results are based on pre-knowledge of the keyframe, then projecting and calculating the respective diameters. The results at this stage for the Sinotubular Junction is R2=0.966 and the Sinus of Valsalva is R2=0.978.
The competing, U-LanD based network by comparison achieved R2=0.66:
We are very pleased with the quality of the data collected and the results of the first iteration so far and are confident we can improve even further.