Acne Scar Classification Using Enhanced CNN and Particle Swarm Optimization

The five types of acne scars rolling, ice pick, boxcar, hypertrophic, and keloidal present significant psychological and skin-related issues. With a global acne occurrence of 9.4 %, the manual diagnosis by dermatologists remains subjective, labor-intensive, and inaccessible in resource-limited settings, highlighting the urgent requirement for an automated solution. To enhance hyperparameter tuning (learning rate, filter size, dropout), this research merges Convolutional Neural Networks (CNN) with Particle Swarm Optimization (PSO) to develop ScarNet, a refined deep-learning framework for accurate acne scar classification. To improve generalization, a newly curated dataset of 250 images from five types of scars underwent preprocessing through Lab color space transformation and guided filtering, followed by augmentation with flipping and rotation. The ScarNet architecture incorporated convolutional layers, batch normalization, and SoftMax classification. The application of PSO optimization significantly boosted performance, surpassing the standalone ScarNet (0.9253) and traditional models such as VGG-16 (0.8900) and Inception-V3 (0.8800), achieving 0.9670 accuracy & Validation Accuracy 0.9799, 0.8521 precision, and 0.8700 recall. These results demonstrate the effectiveness of integrating CNN with PSO in delivering high-quality performance. Performances and adaptable assessments make it a possibly valuable resource for telehealth and medical applications.