Hierarchical Deep Learning Framework for Skin Disease and Cancer Classification Performance Enhancement

Currently, the number of people who have been investigated for skin cancer has increased significantly worldwide. For prior diagnosis, dermatologists can typically visually inspect skin lesions for abnormalities. However, an expert is required, and the similarity of some skin lesions remains challenging. This study aimed to address the challenge of classifying multiple images of skin conditions, including both Benign and Malignant groups, using the hierarchical method. Instead of directly performing multi-class classification using a single model, multiple binary classification models were organized to reduce task complexity and improve overall performance. In the methodology, four convolutional neural network (CNN) models, namely MobileNetV2, EfficientNet-B0, ResNet-18, and ResNet-50, were selected as candidates for this problem. The proposed hierarchical binary classification model was evaluated against conventional multi-class classification methods. As a result, various evaluation metrics were used to assess model performance, with recall as the primary metric in this study, given the emphasis on minimizing false negatives. However, some results revealed discrepancies between the highest recall and other performance metrics. Further analysis demonstrated the potential of using recall as a selection criterion for identifying the most suitable CNN models. The single model-based classification of six classes of skin lesion images achieves the highest recall of 60.27% with MobileNetV2. Meanwhile, the proposed hierarchical model achieves a higher recall of 82.62%, representing a significant increase of 22.35%. Additionally, improvements were observed across all other evaluation metrics, including accuracy (+25.46%), precision (+17.21%), F1-score (+21.34%), balanced accuracy (+12.69%), specificity (+3.03%), and G-mean (+14.25%). These improvements indicate enhanced performance in correctly identifying both positive and negative cases, while reducing misclassification rates. This outcome demonstrates the potential to improve the model’s generalizability, thereby increasing its applicability across various clinical decision-support systems.