Attention-aware Deep Learning Models for Dermoscopic Image Classification for Skin Disease Diagnosis

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin lesions present diagnostic challenges due to visual similarities.
• Early identification and treatment of skin diseases are crucial to prevent severe health issues.

Purpose of the Study:

• To develop and evaluate deep learning models for accurate skin lesion classification using dermoscopic images.
• To enhance model performance by integrating attention mechanisms for focused feature extraction.

Main Methods:

• Utilized four pre-trained Convolutional Neural Network (CNN) architectures: RegNetX, EfficientNetB3, VGG19, and ResNet-152.
• Integrated channel-wise and spatial attention mechanisms into CNNs to improve focus on relevant image regions.
• Optimized model hyperparameters using Bayesian optimization for enhanced performance.

Main Results:

• RegNetX demonstrated superior performance, achieving an accuracy of 98.61% in classifying seven types of skin diseases.
• The integration of attention mechanisms significantly improved the models' ability to identify critical features in dermoscopic images.
• RegNetX with attention mechanisms showed robust performance, highlighting its effectiveness in diagnostic tasks.

Conclusions:

• Attention-aware deep learning models effectively classify skin diseases from dermoscopic images.
• The RegNetX model, enhanced with optimized attention mechanisms, offers accurate and robust diagnoses.
• This technology is critical for the early detection and treatment of various skin conditions.