Skin cancer is a major cause of mortality, and early detection is vital for effective treatment. Diagnosis is challenging because of lesion variability. This study adapts VINCE‐NET, a hybrid deep‐learning model originally designed for stroke detection, to classify melanoma using dermoscopic images.

VINCE‐NET combines vision transformer layers for global context, convolutional neural networks for local features, and long short‐term memory for spatial sequence modeling. During preprocessing, Gaussian blur, normalization, and augmentation were applied to reduce noise and handle class imbalance. During training, the public HAM10000 dataset was used in a central processing unit‐only Google Colab environment (12.72 GB random access memory, 107.7 GB disk) with an AdamW optimizer, a batch size of 12, learning‐rate scheduling, and early stopping (patience = 50). VINCE‐NET's performance was compared with those of a convolutional neural networks, long short‐term memory, residual network with 50 layers (ResNet‐50), visual geometry group network with 16 and 19 layers (VGG‐16/19), and densely connected convolutional network with 121 and 201 layers (DenseNet‐121/201) under identical preprocessing conditions.

VINCE‐NET achieved 94.1% accuracy, 95.5% precision, 90.4% recall, a 92.9% F1‐score, and an area under the receiver operating characteristic curve of 0.98 at a training time of 34,308.42 s. Benchmarks showed that VINCE‐NET outperformed baselines while being computationally efficient.

VINCE‐NET provides competitive performance for melanoma classification and feasibility in resource‐limited settings. Although promising, VINCE‐NET has not been clinically validated yet. Future work will address resolution, ablation studies, interpretability, and external validation.