Quantifying Fractal-Based Features in Dermoscopic Images for Skin Cancer Characterization

Abstract

Accurate skin cancer characterization is crucial for devising effective treatment plans and ensuring optimal patient care. Although dermoscopy has proven invaluable for visualizing skin lesions, accurately determining specific phases or stages based solely on dermoscopy images remains a formidable challenge. In this research, we introduce a novel approach to skin cancer characterization, leveraging the quantification of fractal-based attributes derived from dermoscopic images. Fractal analysis provides a robust framework for capturing the intricate complexity and self-resemblance inherent in a wide array of natural and man-made structures. We harness this methodology to scrutinize the fractal attributes present in dermoscopy images, aiming to unveil distinctive patterns that correspond to different stages of skin cancer. We utilized the box-counting method to extract meaningful features that encapsulate the self-similar characteristics exhibited by skin lesions. To gauge the effectiveness of our approach, we employed an extensive dataset consisting of dermoscopy images portraying lesions in diverse stages of skin cancer. Dermatologists meticulously annotated these images, providing definitive reference information for our comparative analysis. To uncover meaningful patterns and correlations between the extracted fractal attributes and the established stages of skin cancer, we employed a wide spectrum of machine-learning techniques. These encompassed Decision Trees, Logistic Regression, Support Vector Machines, Random Forests, and Convolutional Neural Networks (CNNs). Our results show that the CNN model has the greatest accuracy of 0.77 when categorizing the fractal dimension of the input photos as a feature. We also increased the model's accuracy to 0.85 by utilizing a CNN multi-input approach. This method successfully combines image data with quantified fractal characteristics, resulting in better classification performance. While we acknowledge the difficulty of precisely defining phases merely from dermoscopy pictures, our technique offers dermatologists an additional tool to aid in their clinical decision-making. Our findings contribute to a better understanding of the possible relationships between fractal-based characteristics and skin cancer stages, opening the door for more study and the development of more comprehensive diagnostic tools. These improvements have the potential to increase dermatologists' ability to make enlightened assessments, resulting in better patient outcomes and individualized treatment methods.