Performance Comparison of CNN and ResNet50 for Skin Cancer Classification Using U-Net Segmented Images

Abstract

Skin cancer is a significant global health issue, with melanoma, basal cell carcinoma, and actinic keratosis being the most common types. Early and accurate detection is critical to improve survival rates and treatment outcomes. This study evaluates the performance of Convolutional Neural Networks (CNN) and ResNet50 in classifying segmented images of skin lesions. The dataset, sourced from Kaggle, was pre-processed using U-Net for lesion segmentation to enhance the quality of input data. Both models were trained and evaluated using accuracy, precision, recall, and F1-score metrics. The CNN model demonstrated a balanced performance across classes, with a weighted F1-score of 47%, but suffered from overfitting, as indicated by the divergence between training and validation losses. ResNet50 achieved better recall for basal cell carcinoma (100%) but failed to classify actinic keratosis and melanoma, resulting in a macro F1-score of 23%. The findings reveal that U-Net segmentation improved classification focus but was insufficient to address dataset imbalance and model-specific limitations. This study highlights the challenges of skin cancer classification using deep learning and underscores the importance of addressing data imbalance and overfitting. Future research should explore advanced techniques, such as ensemble methods, data augmentation, and transfer learning, to improve the generalization and clinical applicability of these models. The proposed framework serves as a foundation for further investigation into automated skin cancer detection systems.