Skin Cancer Classification Through Quantized Color Features and Generative Adversarial Network

Introduction

In recent days there is a significant effort for the invention of Computer-Aided Diagnosis (CAD) device for evaluating various stages of cancer and its associated syndromes. This proactive mentality is certainly useful for immediate diagnosis of skin cancer. The demand for personal care tools has been gradually increased in the current COVID situation in many states like the USA, Europe and SAARC countries. The COVID-19 pandemic raised the question of home confinement, and these cases demand diagnostic tools on which individuals can readily approach. In places like western parts of Europe, South America, Australia and New Zealand this condition is most prone to be detrimental and this deadliest form of skin cancer must be recognized early for proper treatment (Pikkula and Whitman 2020). White-skinned individuals are frequently affected by malignant melanoma for direct and continuous sun exposure. Such affected lesions represent an unusual production of melanocytes cells which influence the melanocyte cells rising the fusion of melanin. Skin lesion images are most often taken by electronic devices from patients. There are various open-source image repositories available for research. The ISIC archive (Codella et al. 2018) has an extensive collection of different skin cancer images. The image set contains noisy pixels for the luminosity issue and also the presence of hair. The black round border may be seen in many images which are not desirable and need to be removed strategically. The CAD need pre-processed images to operate on feature extraction process (Dey et al. 2016).

Skin cancer feature descriptors are used in previous works, including texture, shape features and color features (Nezhadian and Rashidi 2017). The feature vector could comprise texture features, morphology features, shape features, ABCD features (Asymmetry, lesion borders etc.) and color features. Apart from shape and ABCD features, color features are also very important for classification (Ewing et al. 2020). These color features have been focused on shades of brown, tan or black. As it grows with different shades of red, white, blue or purple may also appear. Skin image analysis is possible through experiments with color, Coherence Vectors (CCV) (Sallam et al. 2019), histogram-based color features, color moments and correlation of color (Jia et al. 2019). Screening of skin lesion also could be done with color statistics and features from different color space, like HSV Patil and Dongre (2020). In melanoma skin lesions, it is observed that it contains multiple colors are present in pigment network formed with the color pigments distributed randomly (Alfed et al. 2015). The occurrence of different hues of blue color pigments found in the affected area of skin lesions signifies melanoma. The melanoma lesion typically compiles with a distinct combination of brown. The relationship between colors and distribution of shades of those colors can be effective for evaluation of skin lesions. The need for surveying of these color features is growing in current days as it contains top-level visual features. The visual color features have an impact on CAD-like color asymmetry, spectrum color, cluster information of color, and exposure of color (Mabrouk et al. 2020).