Saksenberg Danny, Mukherjee Sandip, Zafar Mohammad A, Ziganshin Bulat, Elefteriades John A
Yale University School of Medicine, New Haven, CT, USA.
Emerge, Johannesberg, SA, USA.
Heliyon. 2024 Jun 28;10(13):e33858. doi: 10.1016/j.heliyon.2024.e33858. eCollection 2024 Jul 15.
Marfan Syndrome (MFS), a genetic disorder impacting connective tissue, manifests in a wide array of phenotypes which can affect numerous bodily systems, especially the thoracic aorta. The syndrome often presents distinct facial features that potentially allow for diagnostic clinical recognition. Herein, we explore the potential of Artificial Intelligence (AI) in diagnosing Marfan syndrome from ordinary facial images, as assessed by overall accuracy, F1 score, and area under the ROC curve.
This study explores the utilization of Convolutional Neural Networks (CNN) for MFS identification through facial images, offering a novel, non-invasive, automated, and computerized diagnostic approach. The research examines the accuracy of Neural Networks in the diagnosis of Marfan Disease from ordinary on-line facial images. The model was trained on 80 % of 672 facial images (182 Marfan and 490 control). The other 20 % of images were used as the test set.
Overall accuracy was 98.5 % (0 % false positive, 2 % false negative). F1 score was 97 % for Marfan facies and 99 % for non-Marfan facies. Area under the ROC curve was 100 %.
An Artificial Intelligence (AI) program was able to distinguish Marfan from non-Marfan facial images (from ordinary on-line photographs) with an extremely high degree of accuracy. Clinical usefulness of this program is anticipated. However, due to the limited and preliminary nature of this work, this should be viewed as only a pilot study.
马凡综合征（MFS）是一种影响结缔组织的遗传性疾病，表现出广泛的表型，可影响多个身体系统，尤其是胸主动脉。该综合征通常具有独特的面部特征，这可能有助于临床诊断识别。在此，我们通过总体准确率、F1分数和ROC曲线下面积评估了人工智能（AI）从普通面部图像诊断马凡综合征的潜力。
本研究探索利用卷积神经网络（CNN）通过面部图像识别马凡综合征，提供一种新颖、非侵入性、自动化和计算机化的诊断方法。该研究考察了神经网络从普通在线面部图像诊断马凡病的准确性。该模型在672张面部图像的80%（182张马凡综合征患者图像和490张对照图像）上进行训练。其余20%的图像用作测试集。
总体准确率为98.5%（假阳性率为0%，假阴性率为2%）。马凡面容的F1分数为97%，非马凡面容的F1分数为99%。ROC曲线下面积为100%。
一个人工智能（AI）程序能够以极高的准确率从非马凡综合征患者的面部图像（来自普通在线照片）中区分出马凡综合征患者。预计该程序具有临床实用性。然而，由于这项工作的有限性和初步性，应仅将其视为一项试点研究。
