Towards the Development of an Acne-Scar Risk Assessment Tool Using Deep Learning

• Published in: 2022 IEEE International Autumn Meeting on Power, Electronics and Computing (ROPEC) Vol. 6; pp. 1 - 6
• Main Authors: Aguilar, Jordan, Benitez, Diego, Perez, Noel, Estrella-Porter, Jorge, Camacho, Mikaela, Viteri, Maria, Yepez, Paola, Guillerno, Jonathan
• Format: Conference Proceeding
• Language: English
• Published: IEEE 09-11-2022
• Subjects: acne scars; binary classification; Computational modeling; Computer architecture; Convolutional neural networks; Deep learning; Estimation; Image analysis; risk assessment; Training
• ISSN: 2573-0770
• DOI: 10.1109/ROPEC55836.2022.10018763

Early estimation of the risk of having acne-induced scars is crucial for acne sufferers to ensure appropriate treatment and prevention. This paper explores the feasibility of using Convolutional neural networks (CNN) to estimate the risk of developing acne-induced scars based only on image analysis as a complementary tool for diagnosis. A database of acne sufferers whose acne-induced scar risks have been evaluated by specialized dermatologists applying a four-item-Acne-Scar Risk Assessment Tool (4-ASRAT) was used. The training dataset includes images of patients with a low, moderate, and high risk of suffering from acne scarring. The dataset was used to train a custom CNN model architecture for the binary and triple classification problem. Poor performance was achieved for the threefold classification problem, while the best model for the binary classification problem achieved an accuracy value of 93.15% and a loss of 19.45% with 0.931 AUC. Although these initial results for the binary classification problem (risk or no risk of developing acne scars in the future) are promising, much work is still required to improve the performance of the model.