M. Kizilgul*a (Dr), N. Doganb (Mr), H. Bostana (Dr), M. Yapicic (Mr), U. Gula (Mrs), R. Karakişd (Mrs), B. Ucana (Dr), E. Dumana (Mr), H. Dugera (Dr), ö. Akine (Prof), E. Cakala (Prof)

a HSU Diskapi Training and Research Hospital, Endocrinology, Ankara, TURKEY ; b Selçuk University, Faculty of Technology, Computer Engineering Department, Konya, TURKEY ; c Ankara University, Elmadağ Vocational School, Computer Technologies Department, Ankara, TURKEY ; d Sivas Cumhuriyet University, Faculty of Technology, Software Engineering Department, Sivas, TURKEY ; e TOBB ETU, Faculty of Science and Literature, Mathematics Department, Ankara, TURKEY

* muhammedkzgl@gmail.com

Introduction: The delay between the onset of the first symptoms and the diagnosis of acromegaly is approximately 6-10 years. Despite efforts for earlier recognition of acromegaly and improved diagnostic tests, the delay from onset of symptoms to diagnosis of the disease has not changed in the last 3 decades. Early detection of acromegaly may increase the possibility of preventing the irreversible complications. Acromegaly is clearly underdiagnosed in the general population, and with active screening, many previously undetected cases of acromegaly can be found. In this study, it is aimed to automatically recognize acromegaly disease from facial images by using deep learning methods and to facilitate the diagnosis of the disease.
Material and methods: The study included 65 acromegaly (52.04±12.12, 29 males/36 females) patients and 71 healthy individuals (48.47±8.91, 39 males/32 females) as the control group, considering gender and age compatibility (p> 0.05). 47/65 (72%) of the acromegaly patients were in remission and the mean IGF-1 levels were 234.67±133.66. In order to classify acromegaly using facial images, both front and side profile images and video recordings of the participants were obtained. Normalized images are obtained by scaling, aligning, and cropping video frames. In the study, three architectures named ResNet50, DenseNet121, and InceptionV3 were used for the transfer learning-based convolutional neural network (CNN) model developed to classify face images as “Healthy” or “Acromegaly”.

Results: The accuracy values obtained for acromegaly and healthy classes with the ResNet50, DenseNet121, InceptionV3, and ensemble method were calculated as 0.9365, 0.9450, 0.9433, and 0.9883, respectively. The average sensitivity, specificity, precision, and correlation coefficient values calculated for the classes of the ResNet50, DenseNet121, and InceptionV3 models are quite close. On the other hand, the ensemble method outperformed these three CNN architectures and provided the best overall performance in terms of sensitivity, specificity, accuracy, precision, F1, and AUC.

Discussion: According to the results obtained, it was seen that acromegaly disease can be detected from facial images with high-performance values using ensemble based-CNN model. The use of artificial intelligence programs might shorten the long diagnosis time in acromegaly patients.

The author has declared no conflict of interest.