A Deep CNN Transformer Hybrid Model for Skin Lesion Classification of Dermoscopic Images Using Focal Loss.

Area of Science:

• Dermatology
• Computer Vision
• Artificial Intelligence

Background:

• Skin cancers are the most common globally, necessitating advanced diagnostic tools.
• Computer-aided diagnosis (CAD) systems are crucial for early skin lesion detection and classification.
• Existing methods face challenges with large, imbalanced datasets like the ISIC 2018.

Purpose of the Study:

• To develop and evaluate an end-to-end hybrid CNN-Transformer model for skin lesion classification.
• To address the challenge of extreme class imbalance in skin lesion datasets using a focal loss function.
• To improve the accuracy and performance of computer-aided diagnosis for skin cancers.

Main Methods:

• Proposed a hybrid model integrating a Convolutional Neural Network (CNN) for local feature extraction and a Vision Transformer (ViT) for global semantic understanding.
• Employed a focal loss (FL) function to mitigate class imbalance issues within the International Skin Imaging Collaboration (ISIC) 2018 dataset.
• Utilized a multi-layer perceptron (MLP) head for the final classification of skin lesion images.

Main Results:

• The hybrid CNN-Transformer model achieved high performance in skin lesion classification.
• The focal loss strategy significantly improved the model's ability to handle class imbalance.
• The proposed model demonstrated superior performance compared to existing methods on the ISIC 2018 dataset.

Conclusions:

• The developed hybrid CNN-Transformer model with focal loss offers a powerful tool for accurate skin lesion classification.
• This approach shows significant potential for enhancing early skin cancer detection and reducing mortality rates.
• The study highlights the effectiveness of combining CNNs and Transformers with specialized loss functions for medical image analysis.