Face2gene is a promising AI-driven software that exploits a facial image analysis framework, DeepGestalt [6]. This in turn is based on computer vision and deep-learning algorithms to quantify the similarities and differences between various syndromes. A two-dimensional image of the patient is to be uploaded by the physician. Given an image, the face is first detected using a cascaded deep convolutional neural network (DCNN) based methods. The face is further divided and cropped into multiple regions using certain facial landmarks, which are geometrically normalized. Scaling is done for each region compared to a fixed size and converted to grayscale. Specialized DCNNs process the facial regions. It then predicts the probability for each syndrome per region of the face that was initially landmarked. A Gestalt model for syndrome classification is then aggregated. Gestalt refers to the information contained in the facial morphology. Finally, a list of top thirty most likely syndromes is displayed based on a combination of facial gestalt as well as features provided by the user.

Though the computer-based recognition of a genetic syndrome may seem at par with clinical recognition, there are certain challenges in this algorithm, which include limited data, as these rely on comparison of the images with established diagnosis and subtle facial patterns. Also important to address are the ethnic differences that exist not only at the global level but also at regional level. For example, it is not possible to use the same measurements for children from southern and eastern parts of India. As more and more images are being added to the database, the deep learning algorithms are expected to become more robust and specific. Another challenge for face2gene is the ability to recognize a normal face. It is unable to do so because of lack of facial digital data from normal individuals for comparison.

The study by Narayanan, et al. [5] is the first of its kind from India, which paves a path for use of this handy software in the clinics. The diagnostic accuracy in this cohort is encouraging. A 72.5% diagnostic rate in patients with dysmorphism gives immense hope to physicians and geneticists to identify rare genetic syndromes while saving precious time and money. However, most of the syndromic diagnoses have been for the more commonly observed disorders, and its efficiency remains to be seen for the rarer syndromes. Further, the study does not have a representation of chromosomal disorders (other than aneuploidies), which are common causes of dysmorphism. Top ten hits seem to be a very liberal criteria to be considered for a positive diagnosis, as it may not reduce the genetic testing algorithms much. Nevertheless, this study proves a place for this application, when used carefully amalgamating the clinical acumen with technology.

In our experience, this tool has been proven to be valuable with diagnosis being established in approximately one-third of patients using this application. The syndromes that were correctly identified included William syndrome, Angelman syndrome, Rubinstein Taybi, Kauffman oculocerebrofacial syndrome, Nicolaides Barraister, Schwartz Jampel syndrome, Coffin Siris, Osteopathica striata with cranial sclerosis and Weidmann Steinner syndrome. One of the challenges observed is the inability of the software to pick Noonan and Turner syndrome, as well as the low gestalt for ear and limb anomalies.

The tool is useful as it saves not only time for a particular patient and family, but also unnecessary costs incurred due to innumerable tests that are required to achieve the diagnosis. The diagnostic odyssey is averted, many a times. There are many more applications of the facial digital recognition technology, including that of pain assessment in young children, and a potential use in unidentified persons data repositories [7,8]. Perhaps, it is only the start of an era of digital technology assisted clinical practice.

Funding: None; Competing interests: None stated.

1. Hsieh TC, Mensah MA, Pantel JT, Aguilar D, Bar O, Bayat A, et al. PEDIA: prioritization of exome data by image analysis. Genet Med. 2019 Jun 5.doi: 10.1038/s41436-019-0566-2. [Epub ahead of print]

2. Hadj-Rabia S, Schneider H, Navarro E, Klein O, Kirby N, Huttner K, et al. Automatic recognition of the XLHED phenotype from facial images. Am J Med Genet A. 2017;173: 2408-14.

3. Valentine M, Bihm DCJ, Wolf L, Hoyme HE, May PA, Buckley D, et al. Computer-aided recognition of facial attributes for fetal alcohol spectrum disorders. Pediatrics. 2017;140: e20162028.

4. Pantel JT, Zhao M, Mensah MA, Aguilar D, Bar O, Bayat A, et al. Advances in computer-assisted syndrome recognition by the example of inborn errors of metabolism.　J Inherit Metab Dis. 2018;41:533-9.

5. Narayanan DL, Ranganath P, Aggarwal S, Dalal A, Phadke SR, Mandal K. Computer-aided facial analysis in diagnosing dysmorphic syndromes in Indian children. Indian Pediatr. 2019;56:1017-9.

6. Gurovich Y, Hanani Y, Bar O, Nadav G, Fleischer N, Gelbman D, et al. Identifying facial phenotypes of genetic disorders using deep learning. Nat Med. 2019;25:60-4.

7. Sakulchit T, Kuzeljevic B, Goldman RD. Evaluation of digital face recognition technology for pain assessment in young children. Clin J Pain. 2019;35:18-22.