Melanoma detection algorithm based on feature fusion.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Dermatology

Background:

• Computer Aided Diagnosis (CAD) systems for melanoma utilize diverse lesion features.
• Feature integration strategies, such as early fusion, are commonly employed but may have limitations.
• Optimal feature combination methods for melanoma diagnosis require further investigation.

Purpose of the Study:

• To compare the effectiveness of early fusion versus late fusion techniques for combining features in melanoma diagnosis.
• To determine the superior feature integration strategy for Computer Aided Diagnosis (CAD) systems in dermatology.

Main Methods:

• Investigated early fusion (combining features into a single vector) and late fusion (combining outputs of individual feature classifiers).
• Evaluated both methods on the PH2 (single-source) and EDRA (multi-source) datasets.
• Utilized classification performance metrics including sensitivity and specificity.

Main Results:

• Late fusion demonstrated superior performance compared to early fusion across both datasets.
• On the PH2 dataset, late fusion achieved Sensitivity = 98% and Specificity = 90%.
• On the EDRA dataset, late fusion achieved Sensitivity = 83% and Specificity = 76%.

Conclusions:

• Late fusion is a more effective strategy than early fusion for integrating diverse features in CAD systems for melanoma diagnosis.
• The findings suggest that late fusion enhances the classification accuracy and diagnostic performance of melanoma detection systems.