Melanoma is still one of the most aggressive types of skin cancer, for which a correct and timely diagnosis is essential for patient survival. In this work, we propose a new hybrid diagnostic framework called FusionDermX-GigaCascade (FDX-GC), which combines EfficientNet-B4, Swin Transformer, and CatBoost models in a cascaded learning process. The proposed framework benefits from the strengths of convolutional feature extraction, transformer-based contextual representation, and gradient boosting for final classification, making it a combination of complementary approaches of deep vision models and ensemble learning. When applied to dermoscopic image datasets, the proposed FDX-GC framework demonstrated an overall diagnostic accuracy of 96.7%, outperforming the state-ofthe-art single-model approaches. In addition to its diagnostic accuracy, the proposed framework is also focused on interpretability, which is essential for trust in decision-making processes. The proposed framework aligns visual attention maps with clinically relevant lesion areas, providing a reliable and explainable approach for melanoma screening, which has the potential to assist dermatologists in early melanoma detection. The proposed FusionDermX-GigaCascade framework illustrates how the strategic integration of models can be used to improve the diagnostic integrity of dermatological imaging. Due to the incorporation of the advantages of both convolutional detail extraction and the contextual analysis by the transformer together with gradient boosting, the model manages to achieve the balance between accuracy and interpretability. The high degree of accuracy of 96.7% makes the model capable of being a clinical aid instead of just a research tool. Perhaps most importantly, the attention-driven explanations are consistent with clinical knowledge in dermatology, ensuring that the model’s predictions are not only accurate but also interpretable.