Enhancing Skin Cancer Classification using Efficient Net B0-B7 through Convolutional Neural Networks and Transfer...

Area of Science:

• Dermatology
• Computer Vision
• Medical Imaging

Background:

• Skin cancer diagnosis is challenging due to visual variations.
• Convolutional Neural Networks (CNNs), particularly EfficientNets, show promise in classification.
• Existing methods struggle with imbalanced datasets and visual complexity.

Purpose of the Study:

• To develop a specialized preprocessing pipeline for EfficientNet models.
• To enhance diagnostic accuracy for multiclass skin cancer classification.
• To leverage transfer learning for improved performance on imbalanced datasets.

Main Methods:

• Developed a tailored image preprocessing pipeline (scaling, augmentation, artifact removal).
• Utilized EfficientNet B0-B7 models with transfer learning from ImageNet weights.
• Evaluated performance using Precision, Recall, Accuracy, F1 Score, and Confusion Matrices.

Main Results:

• Tailored preprocessing and transfer learning significantly improved classification accuracy.
• EfficientNet-B7 achieved the highest top-1 accuracy (84.4%) and top-5 accuracy (97.1%).
• High accuracy for Benign Kertosis (>87%), but challenges remain for Eczema, warts, and psoriasis classification.

Conclusions:

• EfficientNets, especially EfficientNet-B7, demonstrate high potential for precise dermatological image analysis.
• Transfer learning with ImageNet weights is effective for skin cancer classification.
• The optimized models offer a computationally efficient alternative to existing methods.