Acromegaly facial changes analysis using last generation artificial intelligence methodology: the AcroFace system
Purpose: To describe the development of the AcroFace system, an AI-based system for early detection of acromegaly, based on facial photographs analysis. Methods: Two types of features were explored: (1) the visual/texture of a set of 2D facial images, and (2) geometric information obtained from a re...
| Autores: | , , , , , , , , , , , , , , |
|---|---|
| Tipo de recurso: | artículo |
| Estado: | Versión publicada |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Varias* (Consorci de Biblioteques Universitáries de Catalunya, Centre de Serveis Científics i Acadèmics de Catalunya) |
| Repositorio: | Recercat. Dipósit de la Recerca de Catalunya |
| OAI Identifier: | oai:recercat.cat:2445/225294 |
| Acceso en línea: | https://hdl.handle.net/2445/225294 |
| Access Level: | acceso abierto |
| Palabra clave: | Acromegàlia Intel·ligència artificial Diagnòstic per la imatge Acromegaly Artificial intelligence Diagnostic imaging |
| Sumario: | Purpose: To describe the development of the AcroFace system, an AI-based system for early detection of acromegaly, based on facial photographs analysis. Methods: Two types of features were explored: (1) the visual/texture of a set of 2D facial images, and (2) geometric information obtained from a reconstructed 3D model from a single image. We optimized acromegaly detection by integrating SVM for geometric features and CNNs for visual features, each chosen for their strength in processing distinct data types effectively. This combination enhances overall accuracy by leveraging SVM's capability to manage structured, quantitative data and CNNs' proficiency in interpreting complex image textures, thus providing a comprehensive analysis of both geometric alignment and textural anomalies. ResNet-50, VGG-16, MobileNet, Inception V3, DensNet121 and Xception models were trained with an expert endocrinologist-based score as a ground truth. Results: ResNet-50 model as a feature extractor and Support Vector Regression (SVR) with a linear kernel showed the best performance (accuracy δ1 of 75% and δ3 of 89%), followed by the VGG-16 as a feature extractor and SVR with a linear kernel. Geometric features yield less accurate results than visual ones. The validation cohort showed the following performance: precision 0.90, accuracy 0.93, F1-Score 0.92, sensitivity 0.93 and specificity 0.93. Conclusion: AcroFace system shows a good performance to discriminate acromegaly and non-acromegaly facial traits that may serve for the detection of acromegaly at an early stage as a screening procedure at a population level. |
|---|