Skin-lesion segmentation using boundary-aware segmentation network and classification based on a mixture of...

Area of Science:

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

Background:

• Skin cancer, including melanoma, presents diagnostic challenges due to image quality issues and lesion similarity.
• Deep learning and machine learning offer solutions for early, accurate, and efficient skin lesion detection.
• Existing methods struggle with occlusions, poor contrast, and suboptimal imaging conditions.

Purpose of the Study:

• To develop and evaluate a novel boundary-aware segmentation network (BASNet) for robust skin lesion segmentation.
• To introduce a compact convolutional transformer model (CCTM) for accurate skin lesion classification.
• To improve early diagnosis and clinical workflows for skin cancer detection.

Main Methods:

• BASNet utilizes a U-Net-like prediction module with dense supervision and a hybrid loss function.
• A compact convolutional transformer model (CCTM) was designed using specific hyperparameters for classification.
• The CCTM integrates convolutional and transformer layers for enhanced feature extraction.

Main Results:

• The CCTM model achieved over 98% accuracy on six diverse skin lesion datasets (MED-NODE, PH2, ISIC-2019, ISIC-2020, HAM10000, DermNet).
• BASNet demonstrated robust performance even in suboptimal imaging environments.
• The proposed models address challenges like occlusions and poor image quality.

Conclusions:

• The developed models show significant potential for clinical applications in medical image analysis and disease diagnosis.
• The combination of local feature extraction and global context understanding is key to the model's effectiveness.
• Accurate early diagnosis and efficient clinical workflows can be facilitated by these AI-driven tools.