Real-World Clinical Performance of Deep Learning for Quantification and Segmentation of Biventricular Cardiac Size and Function

In this study of routine clinical cardiac MRIs performed for a typical range of clinical indications, we examined the effectiveness of deep learning (DL) for real-world automated quantitative analysis of cardiac size and biventricular function. We find that automated measurements correlate well with skilled readers. While the variation between DL quantification and experts lie within the range seen between experts, there remain several observed failure modes which may benefit from expert supervision. The combination of DL automation with specialist oversight may reduce the time burden of manual segmentation, improve physician efficiency, and promote technique accessibility.