Skin cancer classification using ghostnet-based CNN models

This thesis presents a deep learning hybrid model for the automated diagnosis of skin cancer from dermoscopic images. The research seeks to enhance diagnostic precision as well as lower computational expenses by combining pretrained CNNs with lightweight GhostNet architectures optimized using SMOTE-ENN data balancing. The first approach utilized six pre-trained CNNs VGG16, DenseNet201, EfficientNetB7, EfficientNetV2L, ConvNeXtBase, and MobileNet on the ISIC 2019 dataset to set a strong baseline with 84.63% accuracy and 78% sensitivity of ConvNeXtBase. But these models were costly to deploy in real time. To overcome this constraint, the second approach utilized GhostNet architectures (V1–V3) with Decoupled Fully Connected (DFC) attention and SMOTE-ENN balancing only on the training data. GhostNet V2 obtained 95% accuracy, 93.9% sensitivity, and AUC of 0.99, outperforming big CNNs with parameter reduction of almost 90%. The experiments confirm that efficient models can achieve dermatologist-level diagnostic performance with high efficiency. The suggested model shows a practical, interpretable, and real-time artificial intelligence solution for teledermatology and skin cancer screening.