Researchers have developed a deep learning model that analyzes hand images to detect acromegaly, according to a multicenter observational study published in The Journal of Clinical Endocrinology & Metabolism. The approach uses non-identifying hand features as a potential screening tool for earlier recognition of the disease.
Acromegaly is caused by excessive secretion of growth hormone, typically from a pituitary adenoma. The disorder leads to gradual physical changes such as enlargement of the hands and feet, facial changes, and other systemic complications. Because these changes develop slowly, diagnosis is often delayed; approximately one-quarter of patients experience diagnostic delays exceeding 10 years.
To evaluate whether artificial intelligence could assist with detection, investigators conducted a nationwide multicenter study involving 15 medical centers in Japan. The dataset included 716 adults – 317 with acromegaly and 399 controls – and a total of 11,480 hand images. Participants were aged 18 years or older and were recruited between December 2023 and December 2024.
Two standardized hand images were obtained from each participant: a dorsal hand image and a clenched-fist image with the thumb positioned externally. These views were selected to capture characteristic physical signs of acromegaly while excluding the palm and fingerprint regions to reduce the presence of identifiable biometric features.
The researchers developed a convolutional neural network using a ResNet-50 architecture. Images were standardized for orientation and size before training, and the dataset was divided by institution into training/validation data from 12 centers and a test dataset from three centers. For each patient, predictions were averaged across four images to generate a final classification result.
In the independent test dataset, the model achieved a sensitivity of 0.89 and specificity of 0.91. Positive predictive value was 0.88 and negative predictive value was 0.93, with an F1-score of 0.89. The area under the receiver operating characteristic curve was 0.96.
The investigators also compared the model’s performance with that of 10 board-certified endocrinologists who evaluated the same hand images. In this comparison, endocrinologists achieved F1-scores ranging from 0.43 to 0.63.
Model visualizations indicated that predictions were based on hand morphology, particularly areas around finger joints and the base of the thumb.
The authors noted several limitations, including the higher proportion of acromegaly cases in the dataset compared with typical clinical populations and the inclusion of only Japanese participants, which may affect generalizability. They also emphasized that the tool is intended to support, rather than replace, clinical evaluation.
The findings suggest that image-based artificial intelligence could assist in identifying physical signs of acromegaly and may help support earlier referral for endocrine evaluation in patients with subtle clinical features.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
