09:00 - 10:50
Wed-Park Suites-F
Park Suites
Poster Session
MACHINE LEARNING IN HIGH-FREQUENCY ULTRASOUND SKIN IMAGING FOR COSMETICS ASSESSMENT
91
Presented by: Mariane Massufero Vergilio
Mariane Massufero Vergilio 1, Samara Flamini Kiihl 2, João Batista Florindo 3, Luan Soares de Freitas 2, Matheus Duzzi Ribeiro 2, Dieine Caroline Martins do Nascimento 2, Laura Moretti Aiello 4, Gislaine Ricci Leonardi 1, 4
1 School of Medical Sciences, University of Campinas (UNICAMP)., Campinas
2 Department of Statistics, Institute of Mathematics, Statistics and Scientific Computing, University of Campinas (UNICAMP)., Campinas
3 Department of Applied Mathematics, Institute of Mathematics, Statistics and Scientific Computing, University of Campinas (UNICAMP)., Campinas
4 School of Pharmaceutical Sciences, University of Campinas (UNICAMP)., Campinas
Introduction: An effective way to assess the in vivo performance of anti-aging cosmetics is by using the high-frequency ultrasound (HFUS) skin image technique, a non-invasive approach that allows for a new level of evaluating the content and organization of collagen fibers, especially in the dermis layer. HFUS shows real-time images of the skin layers, appendages, and skin lesions, and can significantly contribute to advances in skin science. Ultrasound-based skin measurements, such as echogenicity and thickness, are frequently used to evaluate skin conditions. Once manual measurements are operator-dependent and time-consuming, much research is being actively conducted on automated methods. However, the existing automated methods are still not specialized in measurements of the skin and its layers.
Purpose of the paper: To establish whether machine learning could be successfully applied to HFUS skin images and to develop new tools for cosmetic claims assessment. Therefore, a model based on artificial intelligence was trained and tested/validated to automate the acquisition of HFUS image parameters from the skin.
Methods: To predict each target variable (echogenicity of dermis and epidermis, thickness of dermis and epidermis) using the ultrasound images, our approach considered supervised machine learning algorithms. The best-performing model was selected for each variable using an independent validation set. For echogenicity variables, Gradient Boosting Machine (GBM) algorithm and Principal Component Regression (PCR) were chosen for thickness variables. Using the proposed method, a labeled dataset containing 144 ultrasound skin images was used for training, and validation was performed with 40 ultrasound skin images.
Results: Our algorithms effectively predicted echogenicity and thickness variables, showing competitive performance. The dermis echogenicity variable stood out with a median absolute error (MAE) of 0.81 and a root mean squared error (RMSE) of 1.11 between the predicted and the expected value.
Discussion and conclusion: Machine learning algorithms are able to reduce the time and increase the quality of the skin ultrasound analysis method. Techniques from the state-of-the-art in the area can have a significant impact on future skin science and cosmetics research activities, resulting in the development of new strategies for cosmetic claims assessment, and optimal cosmetic formulations.