Skin cancer screening using dermoscopic images remains challenging because malignant and benign lesions can share overlapping visual cues, including irregular borders, heterogeneous pigmentation, and complex textures. In addition, real clinical datasets often contain ambiguous samples, acquisition artifacts, and non-uniform illumination, which may degrade training quality and generalization. Although deep learning has achieved strong performance in dermoscopic lesion classification, its dependence on large annotated datasets, limited interpretability, and computational demands can limit adoption in resource-constrained clinical settings. This paper presents a complete classical machine learning framework for four clinically significant dermoscopic categories: basal cell carcinoma (BCC), melanoma, nevus, and pigmented benign keratosis (PBK). A unified hand-crafted feature representation is constructed by combining texture descriptors (Histogram of Oriented Gradients and Local Binary Patterns), region geometry and border regularity measures, Hu moment invariants, lesion intensity statistics, and multi-space colour descriptors (RGB/HSV statistics and HSV histograms). To reduce the effect of label noise and hard-to separate samples, a margin-based smart cleaning strategy removes low-reliability instances on a per-class basis using a preliminary classifier’s confidence margin. Class imbalance is then mitigated through controlled upsampling to a common class count without synthesizing new image content. Feature dimensionality is reduced using minimum-redundancy maximum-relevance (mRMR) ranking, retaining the top 300 features to balance accuracy and runtime. We evaluate both flat multi-class classification and a clinically motivated two-level hierarchical design that first separates Cancer vs non-Cancer, then performs subtype classification within each branch (BCC vs melanoma; nevus vs PBK). Experiments are implemented in MATLAB R2022 using stratified 5-fold cross-validation with strict fold-wise isolation to prevent information leakage. The best flat model, SVM with an RBF kernel, achieves a mean accuracy of 82.87%. The hierarchical system achieves an overall accuracy of 81.80%, with Level-1 accuracy of 86.54%. For Level-1 malignancy detection, ROC AUC = 0.9383 and PR AUC = 0.9367 (all folds combined), indicating strong discrimination. Oracle branch evaluations confirm high intra branch separability (94.04% for BCC vs melanoma; 94.19% for nevus vs PBK) and show that residual loss is primarily due to Level-1 routing errors.
