Hierarchical Two-Phase Skin Lesion Classification Using Reused Convolutional Neural Network Feature Spaces and Cross-Stage Classifier Fusion

Skin cancer, particularly melanoma, represents a major public health concern due to its steadily increasing incidence. Conventional clinical diagnostic methods present several limitations, including inter-observer variability and the risk of misdiagnosis, motivating the adoption of artificial intelligence techniques for the early detection and reliable classification of skin lesions. In this study, we propose a hierarchical multi-level decision classification framework based on progressive deep feature extraction and hierarchical feature fusion. The proposed architecture employs pretrained convolutional neural networks (CNNs) with frozen weights as fixed feature extractors, with the feature representations obtained at each hierarchical level directly reused for the final classification. The processing pipeline consists of several sequential stages. Prior to preprocessing or data augmentation, two evaluation protocols are adopted to assess the robustness of the proposed framework: a conventional 80–20 train–test split and a five-fold cross-validation strategy. The training data are subsequently balanced using a combination of undersampling and oversampling techniques to mitigate class imbalance, while controlled undersampling may also be applied to highly overrepresented classes in the test set. Importantly, the testing sets consist exclusively of original, non-augmented images. The images are then preprocessed using histogram equalization and resized to a resolution of 224×224 pixels. Regions of interest (ROIs) corresponding to skin lesions are subsequently extracted, followed by deep feature extraction using pretrained CNNs. Finally, the global and specialized feature representations are concatenated to construct an enriched hierarchical representation for the final classification stage. Unlike conventional flat multiclass classification approaches, the proposed framework progressively decomposes the decision space into an initial binary classification stage followed by specialized fine-grained classification. The feature representations extracted at different decision levels are hierarchically reused through feature concatenation, providing complementary global and group-specific information for the final classification stage. This hierarchical strategy improves class separability, reduces ambiguities between visually similar lesion categories, and enhances the overall robustness of the proposed system. Four classifiers—Nu-Support Vector Machine (NuSVM), Linear Discriminant Analysis (LDA), Bernoulli Naive Bayes (BNB), and Logistic Regression (LR)—were evaluated on four publicly available dermoscopic datasets. Three experimental scenarios were investigated: binary classification, direct multiclass classification, and the proposed two-stage hierarchical classification framework. Experimental results demonstrate the superiority of the proposed approach, particularly on the PH2 dataset, where accuracies of 98.43% for binary classification, 87.45% for direct multiclass classification, and 96.00% for hierarchical classification were achieved. Furthermore, the five-fold cross-validation results confirm the robustness and generalization capability of the proposed framework across different datasets and evaluation protocols, highlighting the effectiveness of hierarchical feature reuse and classifier fusion for computer-aided skin lesion diagnosis. The proposed pipeline relies on segmentation masks for region of interest extraction; consequently, classification performance may be influenced by the quality of these masks.