Automatic acromegaly detection using deep learning on hand images: a multicenter observational study

CONTEXT: Acromegaly poses clinical challenges in terms of early diagnosis and intervention. Therefore, the development of novel diagnostic tools is essential. Although artificial intelligence (AI) models based on external appearance have been proposed, privacy concerns have limited their use. OBJECTIVE: To develop a privacy-conscious deep learning model for detecting acromegaly using hand images. METHODS: This nationwide multicenter study enrolled 716 patients (317 with acromegaly and 399 controls) and 11 480 images from 15 Japanese pituitary centers. The inclusion criteria were age ≥18 years and care received at the participating facilities. Hand images focusing on the dorsal and fist sign, excluding the palm/fingerprint regions, were used to develop the model. The data were split into training/validation (12 centers) and test (3 centers) datasets. A ResNet-50-based model was trained using PyTorch with data augmentation and 5-fold cross-validation. For each patient, the predictions were averaged over 4 images. The performance of the model was compared with that of endocrinologists. RESULTS: The model achieved a sensitivity of 0.89, specificity of 0.91, positive predictive value of 0.88, negative predictive value of 0.93, F1-score of 0.89, and an area under the receiver operating characteristic curve of 0.96, outperforming specialists (F1-score range: 0.43-0.63). CONCLUSION: This study highlights the utility of dorsal hand and fist sign as diagnostic clues for acromegaly, which the AI model captured more accurately than endocrinologists. Using this privacy-conscious feature, this model can be deployed in public settings like health checkups. Further validation using larger datasets, including healthy individuals and diverse diseases, is necessary.