A Deep-Learning-Based Method Can Detect Both Common and Rare Genetic Disorders in Fetal Ultrasound

Abstract

A global survey indicates that genetic syndromes affect approximately 8% of the population, but most genetic diagnoses can only be performed after babies are born. Abnormal facial characteristics have been identified in various genetic diseases; however, current facial identification technologies cannot be applied to prenatal diagnosis. We developed Pgds-ResNet, a fully automated prenatal screening algorithm based on deep neural networks, to detect high-risk fetuses affected by a variety of genetic diseases. In screening for Trisomy 21, Trisomy 18, Trisomy 13, and rare genetic diseases, Pgds-ResNet achieved sensitivities of 0.83, 0.92, 0.75, and 0.96, and specificities of 0.94, 0.93, 0.95, and 0.92, respectively. As shown in heatmaps, the abnormalities detected by Pgds-ResNet are consistent with clinical reports. In a comparative experiment, the performance of Pgds-ResNet is comparable to that of experienced sonographers. This fetal genetic screening technology offers an opportunity for early risk assessment and presents a non-invasive, affordable, and complementary method to identify high-risk fetuses affected by genetic diseases. Additionally, it has the capability to screen for certain rare genetic conditions, thereby enhancing the clinic's detection rate.

Keywords: artificial intelligence; deep learning; fetal face; genetic diseases; prenatal diagnosis; ultrasound image.

Figure 1

Overview of the acquisition and pretraining processing of datasets workflow. Flowchart summarizing the acquisition and pretraining processing of ultrasound images used in the training and testing sets of the deep learning algorithms for the automatic screening of genetic diseases.

Figure 2

Overview of this study. Data acquisition, including clinical information and fetal ultrasound images, was performed at Guangzhou Women and Children’s Medical Center. Data preprocessing included distortion, zoom-in, tilt, zoom-out, crop, and other methods to augment the training data set. The training and testing were performed by using fetal ultrasound images to develop a deep learning model named Pgds-ResNet for the screening of genetic diseases. The model performance was assessed by AUROC, sensitivity, specificity, and F1*-score. Sonographers with three levels (junior, attending, and senior) of seniority were invited for the human–AI comparison. In case a genetic disease was detected, the abnormal areas were located by exporting the class activation mapping from the networks.

Figure 3

Evaluation results for Pgds-ResNet. (a) ROC curves for screening the presence of genetic diseases in the fetal ultrasound images. (b) Confusion matrix for screening the presence of genetic diseases in the fetal ultrasound images. (c) Comparative experiment between Pgds-ResNet and human doctors. (d) Confusion matrix: Pgds-ResNet framework performance compared with sonographers. AUC = area under the receiver operating characteristic curve. As shown in the color bar chart, the depth of the colors represents the quantity, with lighter colors indicating a larger quantity and darker colors indicating a smaller quantity. The dotted line represents the ROC curve of a completely random classifier.

Figure 4

The heatmap by Grad-CAM algorithm overlaid on original images (with red regions corresponding to more attention in the heatmap on each row). The figure indicates the original images (A) and heatmaps from Grad-CAM (B). The types of diseases are as follows: (1) Turner syndrome, (2) PDHA1 gene mutation, (3) Trisomy 13, (4) Trisomy 18, (5) Trisomy 21, (6) 17q22 microdeletion, (7) 1q21.1q21.2 microdeletion, (8) Helsmoortel–Van der Aa syndrome, (9) 15q26.1–q26.3 deletion and 20p13 duplication, and (10) 15q11.2q13.1 duplication syndrome. As shown in the color bar chart, the importance values in the graph range from −10 to 15. A higher value indicates a greater importance of the pixel for the classification result. The color red represents a higher level of importance, while blue represents a lower level of importance.