Background Scleroderma (SSc) is a rare autoimmune fibrosing multisystem disease with high rates of morbidity and mortality. Scleroderma is usually diagnosed by rheumatologists and/or dermatologists. However, a delay in disease recognition may occur due to a delay in referral, as internists and family physicians are not familiar with the disease and its features. Even though, the facial features of the disease are characteristic for scleroderma, diagnosis is usually based on other signs and symptoms (e.g., skin thickening on extremities), and confirmed with investigations such as autoantibody tests, HRCT, and endoscopies. Late presentation of SSc pts. to rheumatologists is commonly reported^[1]; patients with dcSSc generally presented to their primary health care Practitioner (HCP) after symptoms had persisted for up to 1 year. We hypothesize that facial features of SSc patients are distinctive and can be detected by a trained AI system after processing a mobile phone picture of SSc patient’s face through Convolutional Neural Networks (CNN). This system could be used by family practioner and internists, aiding them to increase suspicion of SSc and refer patients in a timely manner.

Objectives In a pilot study, we aim to examine the ability of an AI facial recognition system to identify SSc related facial features.

Methods Images of SSc pts were compared to a group of age and sex matched normal faces. Deep Learning (DL) - Artificial Intelligence (AI) algorithms evaluated all the pixels in the facial map and identified their utility in the facial recognition prediction models. Using a transfer learning implementation developed by the Danish Viceron ApS AI company. This core model is well-established and experienced algorithm for AI facial feature recognition, based on > 1 million general public faces for facial feature recognition. AI evaluated multiple layers of mathematical models, either isolated or pooled to generate a predictive model and eliminate unnecessary data. Smoothing and uniformity protocols were established for the obtained through preprocessing for the Convolutional Neural Networks (CNN) (Figure 1).

Results Images of 60 SSc pts from the internet were compared to a group of age and sex matched normal faces. We developed models among the 60 SSc facial images and matched controls that were able to identify SSc distinctive facial features with variable specificity and sensitivity. Multiple AI models were used on the training set which includes the first 40 patients and as a partial validation step on the other 20 patients with an equal baseline group built from normal faces. The best pre-trained model, fine-tuned on the training set, achieved 80-90% accuracy on the respective datasets. A larger data set, may provide better accuracy, allowing for more generalized facial recognition. The latter is planned, given the initial reasonable success from the pilot study.

Conclusion Automated pre-processing and the application of AI algorithms in SSc face identification, gave encouraging pilot results (80-90% accuracy). Further testing to develop a larger protocol to establish the effects of race/ethnicity, sex, age, disease duration, disease activity is warranted.

Reference [1]Oliver Distler et al Factors influencing early referral, early diagnosis and management in patients with diffuse cutaneous systemic sclerosis, Rheumatology, Issue 5, May 2018,

• Download figure
• Open in new tab
• Download powerpoint

Figure 1.

The Visual Geometry Group (VGG)-16 CNN architecture and Inception-V3 architecture represent baseline Deep Learning solutions. Based on the robustness the (VGG)-16 CNN architecture was chosen as a starting point.

Acknowledgements: NIL.

Disclosure of Interests None Declared.