Shifted windowing vision transformer-based skin cancer classification via transfer learning

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin cancer is a global health threat, with early detection crucial for effective treatment and preventing metastasis.
• Limited and imbalanced datasets pose challenges for training accurate skin cancer detection models.
• Existing deep learning models, particularly Convolutional Neural Networks (CNNs), struggle with capturing global image relationships due to their local focus.

Purpose of the Study:

• To introduce an innovative deep learning approach for skin cancer categorization using a vision transformer.
• To address the limitations of CNNs in capturing global image features for skin lesion analysis.
• To improve the accuracy and reliability of automated skin cancer identification systems.

Main Methods:

• Development of a novel vision transformer-based model for skin cancer classification.
• Leveraging the attention mechanism inherent in vision transformers to capture long-range dependencies in skin images.
• Employing transfer learning techniques to optimize model parameters and enhance performance.

Main Results:

• The proposed vision transformer model demonstrated superior performance compared to traditional methods in experimental evaluations.
• The architecture effectively captures global image relationships, overcoming a key limitation of CNNs.
• Results indicate the model's efficacy in skin cancer categorization.

Conclusions:

• The developed vision transformer approach shows significant promise for skin cancer identification.
• This method offers a potentially valuable tool to aid clinicians in early and accurate diagnosis of skin lesions.
• The findings highlight the potential of vision transformers in medical image analysis for oncology.