From Clusters to Scores: Unsupervised Acne Detection and Severity Assessment with Deep Learning

Acne detection and severity scoring remain challenging due to the scarcity of labeled data and inconsistencies caused by varying skin tones and lighting conditions. While recent advancements using deep learning models like CNNs have shown promise, they often rely heavily on large labeled datasets. This dependency limits their applicability, particularly in real-world settings where such data is scarce. Moreover, ensuring accurate classification across diverse image conditions continues to be a significant hurdle. To address these challenges, an unsupervised learning framework is proposed, utilizing the Kaggle Acne dataset of 680 unlabeled images. Our approach combines image normalization through NumPy and OpenCV with feature extraction using a pretrained ResNet model. The extracted features are clustered into five severity levels using K-Means, and clusters are labeled based on acne-related characteristics like redness intensity. This method bypasses the need for labeled data, offering a scalable solution for acne severity scoring that can support further dermatological research and clinical applications.