Abstract
Classifying dermatoscopic images is a complex task due to the visual similarity of lesion types. Manual dermatoscopic analysis requires specialized knowledge and experience and is characterized by variable diagnostic accuracy. Automated classification systems require the evaluation of several architectural approaches to determine optimal performance characteristics. This study evaluates seven convolutional neural network architectures for skin lesion classification. The methodology applies intensity normalization and morphological operations to remove artifacts. All models use pre-trained ImageNet weights with transfer learning implementation. DenseNet121 achieves 86.0% validation accuracy and 0.963 AUC. EfficientNet-B1 demonstrates 83.2% accuracy with an AUC of 0.969. Error analysis reveals classification difficulties between actinic keratosis and basal cell carcinoma classes. The custom model demonstrates lower performance with 40.8% accuracy. The results provide a quantitative comparison of architectural performance for dermatoscopic classification applications.