SED: Swin-Efficient DenseNet with Channel Attention for Enhanced Acne Classification

As the eighth most common skin condition in the world, acne requires prompt and precise diagnosis. This study proposed a novel deep learning model called SED to address issues with global feature extraction, channel attention, and data scarcity in acne classification. Deep Convolutional Generative Adversarial Network (DCGAN) is used to augment the dataset with synthetic images to mitigate data scarcity. The model combines a Swin Transformer for capturing highresolution global features, Efficient Channel Attention (ECA) for enhancing channel-wise representation, and DenseNet121 for robust classification. Evaluated on a balanced acne dataset, SED achieved a test accuracy of 97.83 %, a test loss of 0.0615, and reduced training time (931.46 sec), outperforming models like DenseNet121, MobileNetV2, and Xception. These results demonstrate that SED offers an efficient and accurate solution for acne severity classification, supporting its potential application in clinical and telemedicine environments.