An Attention Fusion of Efficientnetv2 and Convnext Towards Effective Accuracy on Skin Cancer

Abstract

Among malignant diseases, skin cancer is one kind that affect survival rate in the global health. Existing methods used such as machine learning (ML) approaches in which Support Vector Machines (SVM), Random Forest (RF), and shallow neural networks uses handcrafted features from dermo-scopic images, depends more on manual feature engineering of color, texture, and shape descriptors, that limits to capture complex lesion patterns, leads to reduced generalization across diverse datasets. The deep learning models in which Convolutional Neural Networks (CNNs) uses VGG, ResNet, and Inception although includes automatic feature extraction but still have issues like overfitting on limited datasets, sensitivity to imaging artifacts in terms of hair, and illumination variation, has trouble of both fine-grained textures and global contextual patterns. Other models like many single-backbone architectures suffer from more false negatives due to class imbalance and insufficient representation learning. Hence, this scenario demands a hybrid deep learning framework of EfficientNetV2 and ConvNeXt for fusion of multi-scale features, and transformer-inspired convolutional learning. This model uses data preprocessing, which includes color normalization, artifact removal, and lesion-focused region-of-interest (ROIs) to achieve quality. In addition, the techniques used such as imbalance-aware loss functions, data augmentation, and transfer learning would provide more robustness, improve accuracy, and reduce the error rate. Results observed state that hybrid deep learning model ensures better values than existing methods considered.