Deep Learning-based Multiview Facial Identification as a Screening Tool for Cushing Syndrome

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Endocrinology

Background:

• Cushing's syndrome (CS) significantly impacts quality of life and is often diagnosed late due to misdiagnosis as simple obesity.
• Distinctive facial features, such as moon face, are characteristic of CS, suggesting potential for image-based diagnostics.

Purpose of the Study:

• To investigate the efficacy of facial identification algorithms using deep learning for diagnosing Cushing's syndrome.
• To compare the diagnostic performance of AI models against human endocrinologists.

Main Methods:

• Collected multi-view facial images from 42 pairs of CS patients and matched controls for training, and 13 pairs for external validation.
• Trained four deep learning models (DenseNet, ResNet, Swin, ViT) using five-fold cross-validation.
• Evaluated models on the external validation cohort and compared their diagnostic metrics against 18 endocrinologists.

Main Results:

• Four deep learning models demonstrated stable performance in the training cohort.
• In the external validation cohort, models achieved high accuracy (92.3%-100%), sensitivity (84.6%-100%), specificity (100%), and AUC (94.4%-100%).
• Deep learning models outperformed endocrinologists in accuracy and matched or exceeded their sensitivity and specificity.

Conclusions:

• Multi-view facial identification with deep learning presents an accurate and accessible diagnostic method for Cushing's syndrome.
• AI models show potential as a novel diagnostic tool to assist clinicians in identifying CS types, surpassing human expert performance.