An Integrated Ensemble Network Model for Skin Abnormality Detection with Combined Textural Features.

Area of Science:

• Dermatology and Artificial Intelligence
• Computational Pathology
• Medical Imaging Analysis

Background:

• Melanoma is the deadliest form of skin cancer, emphasizing the need for advanced diagnostic tools.
• Early detection significantly improves patient outcomes and survival rates for melanoma.
• Current diagnostic methods can be subjective and require expert interpretation.

Purpose of the Study:

• To develop and evaluate an integrated multi-modal machine learning framework for accurate skin cancer detection.
• To combine deep convolutional neural network features with clinical and metadata for improved diagnostic performance.
• To assess the system's efficacy across multiple diverse skin lesion datasets.

Main Methods:

• An ensemble framework integrating transfer-learned image features, textural information, and patient metadata was developed.
• A custom generator was utilized to combine multi-modal data for skin cancer diagnosis.
• The framework employed a weighted ensemble strategy, trained and validated on HAM10000, BCN20000+MSK, and ISIC2020 datasets.

Main Results:

• The model achieved high sensitivity (up to 94.15%) and specificity (up to 99.24%) across datasets.
• Malignant class accuracy reached up to 94%, surpassing physician recognition rates.
• The weighted voting ensemble strategy demonstrated superior performance compared to existing models.

Conclusions:

• The proposed integrated multi-modal ensemble framework offers a robust and accurate approach to skin cancer detection.
• The AI system's performance suggests its potential as an effective initial diagnostic tool for medical professionals.
• Further integration of AI in dermatology can enhance early melanoma detection and patient management.