Detection of Acromegaly from Facial Images Using Machine Learning: a Comparison with Clinical Experts

Context Substantial diagnostic delay in acromegaly contributes to increased morbidity and mortality. Screening attempts in high-risk groups have yielded few positive cases, underscoring the need for simple and precise pre-screening methods. Objective Machine learning analysis of facial images shows promise for acromegaly detection but requires validation in larger, well-characterized cohorts using robust machine learning frameworks as performed in this study. Methods Facial images from different angles were collected via smartphone from 155 acromegaly patients (79% biochemically controlled) and 153 matched controls at all Swedish university hospitals. Six machine learning models were trained to distinguish acromegaly from control images, including three deep neural networks pre-trained on diverse image datasets (ImageNet models: ResNet50, InceptionV2, and DenseNet121) and one network pre-trained specifically on facial images (FaRL). Model performance was compared to assessment by 12 experienced endocrinologists. Results Diagnostic accuracy of the FaRL-based model was superior to all ImageNet models and matched the accuracy of human experts (area under the receiver operating characteristic curve 0.89 for both) with similar specificity (0.87 vs 0.93) but higher sensitivity (0.82 vs 0.66). Classification agreement between the best machine learning model (FaRL) and human experts was 86% for true negatives and 60% for true positives. Both machine learning models and human experts showed greater sensitivity in identifying acromegaly in male patients. Conclusion A deep learning model pre-trained on facial features (FaRL) can detect acromegaly from standard photographs with accuracy comparable to that of expert endocrinologists. This supports the feasibility of face analysis as a screening tool for acromegaly.