Deep Learning Espoused Imaging Modalities for Skin Cancer Diagnosis: A Review

Noor  Sumaiya; Anooja  Ali

doi:10.18502/fbt.v12i4.19827

Noor Sumaiya Department of Computer Science and Engineering, REVA University, Bangalore, Karnataka, India
Anooja Ali Department of Computer Science and Engineering, REVA University, Bangalore, Karnataka, India

DOI: https://doi.org/10.18502/fbt.v12i4.19827

Keywords: Artificial Neural Networks; Convolutional Neural Networks; Dermatology; Generative Adversarial Network; Malignant; Melanoma; Skin Cancer Detection

Abstract

Purpose: Skin Cancer (SC) is one of the most threatening diseases worldwide. Skin cancer diagnosis is still a challenging task. Recently, Deep Learning (DL) algorithms have demonstrated exceptional performance on many tasks compared to the traditional Machine Learning (ML) methods. Particularly, they have been applied to skin disease diagnosis tasks. The aim is to provide a comprehensive overview of the advancements, challenges, and potential applications in this critical domain of dermatology.

Materials and Methods: The review encompasses a wide range of scholarly articles, research papers, and relevant literature focusing on integrating deep learning techniques in skin cancer diagnosis. Materials include studies that employ various imaging modalities such as dermoscopy, histopathology, and other advanced imaging technologies.

The initial phase involves acquiring images of SC from various patients through primary sources and standardized databases. Subsequently, a thorough data cleaning process is implemented, encompassing noise reduction, resizing, and contrast enhancement. Further refinement occurs through the segmentation of the malignant sections, employing edge-based, region-based, and morphological-based techniques. Feature extraction is followed by deep learning approaches, it enhanced with Federated Learning (FL) that is applied to image classification. Finally, leveraging FL-aided deep learning techniques, the images are categorized as either malignant or non-cancerous.

Results: The metrics include Accuracy (AC%), Specificity (Spe%), Sensitivity (Sen%), and Dice Coefficient (DC%), providing a comprehensive evaluation of the classification performance. Generative Adversarial Network (G-AN) demonstrates the highest accuracy 98.5% among the considered techniques, making it the top-performing neural network architecture for skin cancer classification.

Conclusion: This review was undertaken by pulling data from 90 papers published between the years 2019 and 2023, it provides a thorough statistical analysis. A review of various neural network algorithms for skin cancer identification and classification, despite Generative Adversarial Network, has emerged as the most promising approach, underscoring their potential to revolutionize the accurate early diagnosis of skin cancer. Finally, this survey will be beneficial for SCD researchers.