Deep Learning for Improving the Effectiveness of Routine Prenatal Screening for Major Congenital Heart Diseases

Abstract

Early prenatal screening with an ultrasound (US) can significantly lower newborn mortality caused by congenital heart diseases (CHDs). However, the need for expertise in fetal cardiologists and the high volume of screening cases limit the practically achievable detection rates. Hence, automated prenatal screening to support clinicians is desirable. This paper presents and analyses potential deep learning (DL) techniques to diagnose CHDs in fetal USs. Four convolutional neural network architectures were compared to select the best classifier with satisfactory results. Hence, dense convolutional network (DenseNet) 201 architecture was selected for the classification of seven CHDs, such as ventricular septal defect, atrial septal defect, atrioventricular septal defect, Ebstein’s anomaly, tetralogy of Fallot, transposition of great arteries, hypoplastic left heart syndrome, and a normal control. The sensitivity, specificity, and accuracy of the DenseNet201 model were 100%, 100%, and 100%, respectively, for the intra-patient scenario and 99%, 97%, and 98%, respectively, for the inter-patient scenario. We used the intra-patient DL prediction model to validate our proposed model against the prediction results of three expert fetal cardiologists. The proposed model produces a satisfactory result, which means that our model can support expert fetal cardiologists to interpret the decision to improve CHD diagnostics. This work represents a step toward the goal of assisting front-line sonographers with CHD diagnoses at the population level.

Explainable Artificial Intelligence and Cardiac Imaging: Toward More Interpretable Models

“…For instance, Pérez-Pelegrıó et al 28 developed a new explainable approach that combines class activation mapping with U-net to automatically estimate the LV volume in end diastole and obtain the result in the form of a segmentation mask without segmentation labels to train the algorithm. Grad-CAM was used in 7 cardiac imaging studies, either for classification 18,34,40,[48][49][50] or segmentation. 51 The latter in particular proposed a new interpretable CNN model (fast and accurate echocardiographic automatic segmentation based on U-Net) that integrates U-net architecture and transfer learning (from Visual Geometry Group 19) to segment 2-dimensional echocardiography of 88 patients into 3 regions (LV, interventricular septal, and posterior LV wall).…”

“…Grad-CAM was used in 7 cardiac imaging studies, either for classification 18,34,40,48–50 or segmentation. 51 The latter in particular proposed a new interpretable CNN model (fast and accurate echocardiographic automatic segmentation based on U-Net) that integrates U-net architecture and transfer learning (from Visual Geometry Group 19) to segment 2-dimensional echocardiography of 88 patients into 3 regions (LV, interventricular septal, and posterior LV wall).…”

“…The aims of these articles were diverse and included the automatic calculation of the LV volume and ejection fraction 28,35 ; the automatic segmentation of cardiac structures as LV/RV, myocardium, interventricular septal, and posterior LV wall 39,51,55 ; and classification approaches. Classification was investigated in the context of cardiac disorders such as mitral valve diseases, 48 myocardial injury, 50 coronary artery disease (CAD), 34,54 cardiomyopathy, 33 congenital heart disease, 40 and HF 32 ; even in newborn 49 for the presence of intracardiac devices (eg, catheters, pacemaker, and defibrillator leads) or motion 18 ; for severe left atrial dilation and LV hypertrophy 43 ; and for CMR image view. 27 In order to improve the interpretability of the segmentation/prediction/classification results, these articles generally focused on DL-based XAI methods as class activation mapping or Grad-CAM, SmoothGrad, saliency maps, testing with concept activation vectors, and guided backpropagation.…”

Evolving the Era of 5D Ultrasound? A Systematic Literature Review on the Applications for Artificial Intelligence Ultrasound Imaging in Obstetrics and Gynecology

“…Dozen et al presented a method specific to the interventricular septum [55]; Han et al focused on the assessment of the left ventricle and left atrium [56]; and Xu et al aimed at identifying seven anatomical structures, namely right and left atrium and ventricle, thorax, descending aorta, and epicardium [57]. The automated detection of standard views was investigated by Wu et al, Yang et al,, while CHD was effectively detected by AI models generated by the research groups of Gong et al and Nurmaini et al [61,62], and selectively for the diagnosis of total anomalous pulmonary venous connection by Wang et al [63].…”

“…The complexity of fetal echocardiography itself is derived from the skill needed to detect even the smallest anatomical abnormalities in a beating organ, which makes it an interesting and challenging research area for AI applications. Advantages are the facilitation of standard view acquisition [58,60] and CHD detection [53,61,62], as well as a significant reduction in examination time [52,55]. Furthermore, Arnaout et al outlined the benefits of their AI model for telehealth approaches and diagnoses of rare diseases [50].…”