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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Genetics

Background:

• Genetic syndromes impact 8% of the population, with diagnoses typically occurring post-birth.
• Current facial identification technology is not suitable for prenatal diagnosis.
• Abnormal facial features are indicators of various genetic diseases.

Purpose of the Study:

• To develop an automated prenatal screening algorithm for detecting high-risk fetuses with genetic diseases.
• To assess the efficacy of the algorithm in identifying specific chromosomal abnormalities and rare genetic conditions.

Main Methods:

• Development of Pgds-ResNet, a deep neural network-based algorithm for automated prenatal screening.
• Utilizing facial feature analysis for the detection of genetic abnormalities in fetuses.
• Comparative analysis against experienced sonographers' performance.

Main Results:

• Pgds-ResNet demonstrated high sensitivity and specificity in screening for Trisomy 21 (0.83/0.94), Trisomy 18 (0.92/0.93), Trisomy 13 (0.75/0.95), and rare genetic diseases (0.96/0.92).
• Algorithm's detected abnormalities, visualized via heatmaps, align with clinical findings.
• Performance comparable to experienced sonographers in comparative experiments.

Conclusions:

• Pgds-ResNet provides a non-invasive, affordable, and complementary method for early prenatal risk assessment of genetic diseases.
• The algorithm enhances detection rates for various genetic conditions, including rare diseases.
• This technology offers a significant advancement in prenatal screening and diagnosis.