Artificial intelligence-based diagnosis in fetal pathology using external ear shapes.

Area of Science:

• Medical Imaging
• Computational Biology
• Genetics

Background:

• Syndromic ear malformations are key indicators of genetic disorders.
• Accurate prenatal diagnosis of these syndromes is crucial for timely intervention.
• Existing diagnostic methods can be limited in scope or accessibility.

Purpose of the Study:

• To develop and validate an automated tool for fetal ear phenotyping.
• To distinguish between CHARGE syndrome, Mandibulo-Facial Dysostosis Guion Almeida type (MFDGA), and control fetuses based on ear morphology.
• To assess the performance of machine learning in classifying syndromic ear phenotypes.

Main Methods:

• Training a machine learning model on a large dataset of children's ear photographs.
• Utilizing geometric morphometry for landmark extraction from ear images.
• Applying the trained model to classify fetal ear photographs from control and syndromic groups.

Main Results:

• The automated phenotyping tool achieved an overall accuracy of 72.6% in classifying fetal ears.
• Specific accuracies for CHARGE, control, and MFDGA fetuses were 76.4%, 74.9%, and 86.2%, respectively.
• Area under the curve values ranged from 86.8% to 90.3%, indicating good discriminative ability.

Conclusions:

• The study presents the first automated fetal ear phenotyping model.
• The model demonstrates satisfactory classification performance for CHARGE and MFDGA syndromes.
• Further validation is necessary before clinical implementation as a diagnostic aid.