Title : Automated assessment of atrial septal defects using deep learning
Abstract:
Aim: Atrial septal defect (ASD) is a typical cardiac defect in the atrial septum, which accounts for around 10% of congenital heart disease (CHD). The early diagnosis and treatment of ASD are crucial to avoid serious complications. Conventional pre-treatment assessment methods to measure critical scales of ASD is transthoracic echocardiogram (TTE). But in light of the challenges faced by the current Chinese pediatric healthcare system, including generally unmet demand and uneven development, a significant number of inadequately trained pediatricians working in pediatrics at primary hospitals. Thus, we aimed to develop an automated and interpretable assistant for TTE-based assessment of ASD using deep learning (DL).
Methods: We created a novel Deep Keypoint Stadiometry (DKS) model designed to precisely localize the keypoints, representing the endpoints of ASDs, followed by absolute distance measurement with scale information. Clinical decision rules were applied to derive closure plans and determine the size of the ASD occluder for transcatheter closure. We retrospectively collected a dataset of 3,474 2D and Doppler TTE scans from 579 patients across two clinical centers.
Results: The DKS model demonstrated a high accuracy of closure classification (0.9425±0.0052) in within-center evaluations. Consistent results were obtained in cross-center cases and using the quadratic weighted kappa as an evaluation metric. The fine-grained keypoint labels provided explicit supervision for network training. Clinicians can intervene and make edits at various stages of the automated process.
Conclusions: A transparent AI-based multi-view echocardiogram analysis system is proposed to suggest transcatheter or surgical closure. Our DKS model provides interpretable and accurate AI-assisted suggestions. In the future, similar studies using deep learning algorithms shall be developed for size-sensitive treatments. By identifying the explicit clinical practice guidelines, our deep keypoint stadiometry algorithms were able to automatically propose the therapeutic plan, effectively reducing the workload of the clinicians.
