From traditional to deep learning methods for skin lesion segmentation: A literature review

Abstract

Due to their similarity, skin lesions are challenging for early and precise diagnosis, and segmenting them manually is a laborious and time-consuming task. Therefore, several methods, including traditional and artificial intelligence methods, have been proposed for segmenting the skin lesions automatically. The majority of recent reviews, however, have overlooked the traditional methods and instead only briefly discussed their advantages and disadvantages. Consequently, this review aims to provide a thorough overview of this field and offer insights about the role and performance of the traditional methods in the era of artificial intelligence. We collected 128 peer-reviewed and high-quality publications using inclusion and exclusion criteria. These publications have been selected from four well-known databases, IEEE Xplore, ScienceDirect, Springer, and MDPI, and classified systematically into three categories: traditional, artificial intelligence, and hybrid methods. The main contributions of this paper are: (1) Exploring the most used traditional methods and performance enhancement techniques. (2) Providing the State of the art of the convolutional neural network and deep learning methods. (3) Highlighting the role of traditional methods in hybrid frameworks. (4) Providing an extensive performance analysis and discussing the computational cost and real-world applicability challenges associated with segmentation methods. The results highlight the superiority of deep learning methods over traditional methods, while also recognizing the relevance of traditional methods like Active contours and clustering as trusted tools for skin lesion segmentation. These methods benefit from improvements such as ensembling techniques and optimization algorithms. Additionally, 9% of reviewed papers were motivated by the combination of benefits from traditional methods and deep learning methods to develop more robust models. In conclusion, significant challenges remain in developing hybrid end-to-end trainable models, and models optimized for artificial intelligence-powered devices, suitable for real-world applications. Following our findings, we propose several recommendations to address data representativity, annotation variability, and model debiasing.