That’s why we developed EFFACLAR SPOTSCAN. This revolutionary technology is the result of two years work developing an artificial intelligence (A.I.) algorithm able to evaluate your skin and provide you a recommended routine. The result is the first personalized acne diagnosis tool. It has been developed with dermatologists to provide the best solutions for your skin.
To build the most accurate and innovative algorithm, La Roche Posay brought together dermatologists from around the world to analyze data from men and women with different skin types, ethnicities and clarity of skin. Sophie Seité, International Scientific Director, La Roche Posay takes us through the process of building this intelligent and accurate algorithm.
Step 1: Collect data
We collected more than 6000 pictures of men and women, representing all skin phototypes. Each patient had 3 photos taken - from the front and each side of the face - with iOS and android smart phones.
Step 2: Clean, prepare and manipulate data
Once the pictures were gathered, we contacted three expert dermatologists in order to assess for each set of three pictures, the severity of patient acne based on the GEA grade. This scale represents six different grades. The grade chosen for each patient was the one given by the majority of the dermatologists.
Furthermore we requested a dermatologist to locate and quantify the total number of inflammatory lesions – papules or pulse Jules – non-inflammatory lesions – open or closed, downs – and pigmented marks found on different patient pictures.
Step 3: Train the algorithm model
All this information was sent to the developing engineers to program the artificial intelligence algorithm.
Step 4: Test the data is accurate
In order to scientifically verify the accuracy of the algorithm, we carried out a clinical study on 53 patients, supervised by three dermatologists. This study allowed us to evaluate different parameters regarding the severity grades. When evaluating the patients face-to-face, dermatologists agreed on the severity grade on 67% of cases. When evaluating the above mentioned patient pictures, dermatologists agreed on the severity grade of on 59% of cases. Our algorithm matched the dermatologist evaluation on 48% of cases. This accuracy level was too low, as our objective was to reach a 60% accuracy to approach the dermatologist accuracy. Accordingly, we decided to keep improving our algorithm.
Step 5: Improve the algorithm
After several versions of our algorithm, we managed to reach an accuracy level of 68% for the GEA grade, which evaluates the severity of acne. The result was better than the objective of 60% we had originally set. Additionally, we have worked on the algorithm to be able not only to quantify but also to locate the acne lesions and pigmented marks on the patient’s face with a high level of accuracy. Thanks to statistical tests we have performed, we know that our algorithm has 84% Accuracy on inflammatory lesions, 72% accuracy on pigmented marks, 61% accuracy on non-inflammatory lesions.
We plan to publish these excellent results and present them during a symposium at the next world Congress of dermatology.