Detection of Acromegaly From Facial Images Using Machine Learning: A Comparison With Clinical Experts

Area of Science:

• Endocrinology
• Medical Imaging
• Artificial Intelligence

Background:

• Diagnostic delays in acromegaly increase morbidity and mortality.
• Current screening methods in high-risk groups are inefficient.
• There is a need for simple, precise prescreening tools for acromegaly.

Purpose of the Study:

• To evaluate the efficacy of machine learning models in detecting acromegaly from facial images.
• To compare the performance of different deep learning models against human expert diagnosis.

Main Methods:

• Facial images were collected from 155 acromegaly patients and 153 controls using smartphones.
• Six machine learning models, including deep neural networks (ResNet50, InceptionV2, DenseNet121, FaRL), were trained.
• Model performance was benchmarked against diagnoses made by 12 experienced endocrinologists.

Main Results:

• The FaRL model achieved an area under the receiver operating characteristic curve of 0.89, matching human experts.
• FaRL demonstrated higher sensitivity (0.82) compared to ImageNet models and human experts (0.66).
• Classification agreement between FaRL and experts was 86% for true negatives and 60% for true positives.

Conclusions:

• A deep learning model (FaRL) pretrained on facial features can detect acromegaly from photographs with expert-level accuracy.
• Facial analysis using AI presents a feasible and effective screening tool for acromegaly.
• This technology could help reduce diagnostic delays and improve patient outcomes.