DSSCC net enhanced skin cancer classification using SMOTE Tomek and optimized convolutional neural network

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin cancer diagnosis relies on expert evaluation, which can be error-prone.
• Current deep learning models struggle with imbalanced datasets, hindering accurate classification of rare skin lesions.
• There is a need for robust, efficient, and interpretable AI models for early skin cancer detection.

Purpose of the Study:

• To develop and evaluate DSSCC-Net, a novel deep learning framework for accurate skin cancer classification.
• To address the challenge of class imbalance in dermoscopic datasets using SMOTE-Tomek oversampling.
• To enhance the interpretability of the classification model for clinical application.

Main Methods:

• Developed DSSCC-Net, a deep convolutional neural network (CNN) integrated with SMOTE-Tomek oversampling.
• Trained and validated the model on diverse dermoscopic datasets (HAM10000, ISIC 2018, PH2).
• Utilized data augmentation, dropout layers, ReLU activation, and Grad-CAM for model optimization and interpretability.

Main Results:

• DSSCC-Net achieved high classification accuracy (average 97.82% ± 0.37%) and AUC (99.43%).
• The model demonstrated superior performance compared to existing state-of-the-art models like VGG-16 and ResNet-152.
• SMOTE-Tomek integration significantly improved minority-class detection, with balanced precision (97%) and recall (97%).

Conclusions:

• DSSCC-Net offers a robust and interpretable solution for skin cancer classification, outperforming current methods.
• The model's ability to handle class imbalance makes it suitable for real-world clinical deployment.
• DSSCC-Net sets a new benchmark, paving the way for AI-assisted early skin cancer diagnosis.