Endocrine Abstracts

Objective: Even with advancements in diagnostic techniques, acromegaly continues to be diagnosed late. This study aimed to employ deep learning methods to automatically identify acromegaly disease from images of faces, hands, and feet to enhance the efficiency of disease detection.

Design: Cross-sectionaly single-center study with deep learning

Methods: The study included 71 acromegaly patients and 65 healthy controls. All patients had images of their faces and hands; images of the feet were available for 63 acromegaly patients and 48 control patients. The images were processed with TensorFlow to increase the sample size through various transformations such as black and white transformationy rotation, zoom, and shearing. The Convolutional Neural Network (CNN) algorithm was employed across 8 layers in the neural network model. ’Adam’ was utilized as the optimizer, and the binary cross-entropy loss function was employed. Early stopping was implemented to monitor the loss.

Results: The average age of the acromegaly group was 46.9 (SD:11.2), while the control group had an average age of 45.2 (SD:13.5). There were 33 males 46.5% in the acromegaly group and 25 males 39.1% in the control group. When the gender distribution and age averages were examined between the groups using the chi-square test and Student t-test, no significant differences were found. In the acromegaly group, the average duration of the disease was 4.12 (SD:4.88), ears, with 28 (39.4%) of them being in remission. The algorithm achieved an accuracy of 69.37% on facial photos, with a sensitivity of 57.14% and specifiCity of 85.42%. For hand photos, the accuracy was 92.79%, sensitivity was 92.06%, and specifiCity was 93.75%. In the case of foot photos, the algorithm demonstrated an accuracy of 70.27%, sensitivity of 49.21%, and specifiCity of 97.92%.

Conclusions: The CNN-based neural network exhibited varying accuracies across facialy hand, and foot photos, with notable accuracy for hands and moderate accuracy for facial photos. However, its performance on foot images was less robust. Incorporating the Random Forest ensemble method significantly boosted the overall accuracy to 94.59%, emphasizing the effectiveness of combining these algorithms for improved performance across diverse image datasets. The study validated the potential of the deep learning model in detecting acromegaly from facial images, highlighting the potential role of artificial intelligence programs in future acromegaly diagnosis.