EFFNet: A skin cancer classification model based on feature fusion and random forests

Area of Science:

• Dermatology
• Computer Science
• Artificial Intelligence

Background:

• Deep learning models for skin cancer classification face challenges like imbalanced datasets and feature redundancy.
• Existing methods often overlook feature interactions across convolutional layers, limiting diagnostic accuracy.

Purpose of the Study:

• To develop an advanced computer-aided diagnosis model, EFFNet, for improved skin cancer classification.
• To overcome limitations of current deep learning approaches in handling complex skin lesion data.

Main Methods:

• The EFFNet model utilizes image enhancement for dataset balancing and fine-tunes EfficientNetV2 with pre-trained weights.
• An improved hierarchical bilinear pooling method captures inter-layer feature interactions.
• Random forests are employed for the final classification prediction.

Main Results:

• EFFNet achieved high performance metrics: 94.96% accuracy, 93.74% recall, 93.16% precision, and 93.24% F1-score.
• The model demonstrated a significant improvement over existing methods, with accuracy increases up to approximately 10%.

Conclusions:

• EFFNet offers a robust solution for skin cancer classification by effectively managing data imbalance and enhancing feature representation.
• The proposed model shows considerable potential for clinical application in computer-aided dermatological diagnosis.