Integrating color histogram analysis and convolutional neural networks for skin lesion classification

Area of Science:

• Dermatology and Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Skin lesion color is vital for diagnosing melanoma and other skin conditions.
• Existing methods often overlook the quantitative aspect of lesion color diversity.
• A novel approach is needed to leverage color complexity for improved diagnostic accuracy.

Purpose of the Study:

• To introduce and validate the number of colors as a novel diagnostic feature for skin lesions.
• To develop and evaluate a Convolutional Neural Network (CNN) for classifying skin lesions based on color count.
• To assess the clinical potential of color number analysis in melanoma detection.

Main Methods:

• Utilized color histogram analysis on three public datasets (PH2, ISIC2016, Med-Node) to quantify lesion colors.
• Developed a 19-layer CNN with residual blocks, incorporating skip connections for enhanced feature learning.
• Employed DeepDream for visualization and LIME for feature interpretability to understand model decisions.

Main Results:

• The number of colors in skin lesions was confirmed as a significant diagnostic feature.
• The proposed CNN achieved a highest weighted F1-score of 75.00% in classifying lesions by color count.
• CNN configurations with three skip connections demonstrated superior performance.

Conclusions:

• The number of colors present in skin lesions is a valuable metric for disease severity and melanoma differentiation.
• The developed CNN model shows promise for clinical integration in melanoma diagnosis.
• This approach complements traditional diagnostic methods, potentially improving early detection rates.