This study explores the performance of custom Convolutional Neural Network (CNN) architectures for both binary and multiclass skin disease classification, utilizing datasets sourced from Kaggle and Google. Images of ringworm and healthy skin were used, resized to pixels, and augmented with techniques such as flipping and rotation to address data limitations. We experimented with two dataset splits (80:20 and 70:30) and compared our custom CNN’s performance against State-of-the-Art (SOTA) models and Vision Transformers (ViTs). For binary classification, our custom CNN architecture included four convolutional layers (32, 64, and 128 filters in successive blocks) with ReLU activation after each convolutional layer, followed by max-pooling layers and dense layers, and a final softmax output. This model achieved 98.9% accuracy, demonstrating strong performance in distinguishing ringworm from healthy skin. For the multiclass (23-class) classification task, we adapted the CNN architecture with added class frequency-based weights and achieved 35.23% accuracy, illustrating the challenges in multi-class dermatological classification. Our results indicate the practical applicability of a custom CNN, modified to incorporate class frequency balancing, for dermatological contexts, while highlighting the potential and limitations compared to SOTA architectures and ViTs in skin disease classification.