Multi skin lesions classification using fine-tuning and data-augmentation applying NASNet.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Skin lesions are common indicators of various diseases, including skin cancer.
• Climate change and treatment costs necessitate improved skin cancer prevention strategies.
• Computational methods, particularly Convolutional Neural Networks (CNNs), are increasingly used for medical image analysis.

Purpose of the Study:

• To develop an accurate, automated system for classifying eight common skin diseases without segmentation.
• To leverage the NASNet architecture for enhanced performance in skin lesion recognition.
• To compare the proposed CNN model's efficacy against other trained architectures.

Main Methods:

• Utilized a NASNet-based CNN architecture trained end-to-end on augmented skin lesion images from the International Skin Imaging Collaboration (ISIC) dataset.
• Initialized CNN weights from ImageNet and fine-tuned for discriminating between skin lesions.
• Applied 10-fold cross-validation for robust evaluation and calculated accuracy, sensitivity, and specificity.

Main Results:

• The proposed NASNet-based CNN system achieved higher accuracy compared to other evaluated CNN architectures.
• Demonstrated a significant reduction in training parameters, suggesting computational efficiency.
• The model accurately classified eight different types of skin diseases without requiring image segmentation.

Conclusions:

• The NASNet architecture offers a novel and effective approach for multi-class skin disease classification using medical images.
• This system represents a significant advancement in automated skin lesion analysis and early disease detection.
• The findings highlight the potential of deep learning for improving skin cancer prevention and diagnosis.