Addressing Challenges in Skin Cancer Diagnosis: A Convolutional Swin Transformer Approach

Area of Science:

• Dermatology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Skin cancer diagnosis is challenging due to lesion variability and image artifacts.
• Accurate and early detection of skin cancer is critical for patient survival.
• Existing diagnostic methods face limitations in handling complex lesion characteristics.

Purpose of the Study:

• To introduce a novel Convolutional Swin Transformer (CSwinformer) method for precise skin lesion segmentation and classification.
• To address challenges in skin cancer diagnosis, including variations in lesion appearance and image quality.
• To enhance the accuracy and efficiency of automated skin cancer detection systems.

Main Methods:

• A multi-phase framework involving data preprocessing (Gaussian filtering, Z-score normalization, augmentation), segmentation using Swinformer-Net (Swin Transformer and U-Net integration), and classification via MD-CNNFormer (Multi-Scale Dilated Convolutional Neural Network meets Transformer).
• Utilized four benchmark datasets: HAM10000, ISBI 2016, PH2, and Skin Cancer ISIC for comprehensive evaluation.
• Developed and validated a novel deep learning architecture for medical image analysis.

Main Results:

• The CSwinformer method demonstrated superior performance compared to traditional approaches.
• Achieved high metrics: 98.72% classification accuracy, 98.06% pixel accuracy, and 97.67% Dice coefficient.
• The proposed segmentation and classification framework proved effective on diverse datasets.

Conclusions:

• The CSwinformer method offers a promising solution for accurate skin lesion segmentation and classification.
• This AI-driven approach can significantly support clinicians in the early and accurate diagnosis of skin cancer.
• The study highlights the potential of integrated Transformer and CNN models in advancing dermatological diagnostics.