Acne Classification using Deep Learning Models

In this research, different techniques were explored to improve the accuracy of the classification of acne vulgaris into seven classes: whitehead, blackhead, fungal acne, papules, pustules, nodules, and acne fulminans. Three pre-trained deep learning models, viz., Inception-ResNet, MobileNet and Efficient-Net were trained on the Skin90 dataset. Several experiments were conducted on the dataset using these models by utilising certain loss functions and exploring hyperparameters yielding higher values of metrics. Additional techniques employed include traditional data augmentation and augmentation using generative models like Variational Auto-Encoder(VAE). The highest accuracy obtained on the original dataset was 89% by the EfficientNet model with a learning rate of 0.001. The best performance on the VAE-augmented dataset was observed on the EfficientNet model with a learning rate of 0.1 achieving an accuracy of 84%.