Skin diseases are a prevalent health issue, impacting physical and mental well-being. Traditional diagnosis relies on visual inspection and subjective evaluation, leading to potential misdiagnoses, particularly in areas with limited dermatologist access. Artificial intelligence (AI), specifically ML and DL, offers a promising solution. These algorithms can be trained on large datasets of skin images to learn patterns associated with different skin diseases, potentially improving diagnostic accuracy and objectivity. This review examines the evolution and recent achievements of ML and DL in dermatological diagnosis, highlighting current challenges and offering recommendations for future progress. The aim is to explore how improved algorithms can enhance diagnostic accuracy and speed, supporting the development of computer-aided diagnosis (CAD) systems for skin diseases.

The authors conducted a systematic review of literature from PubMed, IEEE, SpringerLink, and Web of Science, focusing on advancements from 2015-2023. They included original research papers on segmentation and classification algorithms for skin lesions using ML or DL, excluding review articles, case reports, and books. The review incorporated 29 articles on dermatological image segmentation and 45 on classification. Commonly used datasets included DermNet, MED-NODE, DermIS, and various ISIC datasets (2017, 2018, 2019, 2020), as well as Derm7pt. These datasets contain thousands of images representing a range of skin conditions.

The review categorized algorithms into traditional ML and DL methods for both segmentation and classification tasks. For segmentation, traditional ML methods included Linear Spectral Clustering (LSC), Random Forests, K-means clustering, morphological image processing, Independent Component Analysis (ICA), and Fuzzy C-Means (FCM). DL methods focused on U-Net, Fully Convolutional Networks (FCNs), and variations such as Fully Convolutional Deconvolution Networks (FCDNs), Deep Fully Convolutional Networks (DFCNs), Convolutional-Deconvolutional Neural Networks (CDNNs), and Multi-Scale Fully Convolutional DenseNets (MSFCNDs). For classification, traditional ML methods included Support Vector Machines (SVMs), logistic regression, Random Forests, K-Nearest Neighbor (KNN), and Naive Bayes. DL methods predominantly used Convolutional Neural Networks (CNNs), such as AlexNet, VGG, GoogleNet, ResNet, Inception-ResNet-V2, MobileNet, EfficientNetV2, and custom hybrid networks combining DenseNet and residual networks. The review also considered the use of transfer learning, adversarial networks, and other innovative techniques. Evaluation metrics included accuracy, sensitivity, specificity, Dice score, and Jaccard index.

The review found that deep learning methods generally outperformed traditional machine learning techniques in both image segmentation and classification tasks. In segmentation, while traditional methods showed some progress, DL methods, particularly those using U-Net architectures and their variants, achieved higher Jaccard indices (average of 0.821 vs. a maximum of 0.81 for traditional methods). In classification, DL methods demonstrated higher and more stable accuracy, sensitivity, and specificity, particularly for neoplastic and inflammatory skin diseases. DL achieved an impressive maximum accuracy of 99.7% in one study. While methods for pigmented diseases like vitiligo also showed promising results (accuracies exceeding 85%), further research is needed to validate generalization and address data limitations. The review highlights several architectures like U-Net and its variants, ResNet, and EfficientNet that showed high performance across different tasks. The incorporation of techniques like transfer learning and adversarial networks also enhanced performance. The analysis also revealed that CNNs are the most prevalent DL architecture used in dermatological image analysis, reflecting its success in image recognition tasks.

The findings demonstrate the significant potential of DL in improving the accuracy and efficiency of skin disease diagnosis. The superior performance of DL over traditional ML methods is attributed to its ability to automatically learn complex features from data, reducing the need for extensive manual feature engineering. However, the challenges related to dataset limitations, the explainability of DL models, the concentrated research focus on melanoma and skin cancer, and the need for more innovative algorithms remain critical. The 'black box' nature of DL models hinders clinical adoption, emphasizing the need for improved interpretability. The overrepresentation of melanoma and inflammatory conditions in the literature also limits the applicability of current models to other skin diseases, including pigmented lesions. Addressing these limitations is crucial for realizing the full potential of AI in dermatological practice.

Deep learning methods show significant promise for improving skin disease diagnosis, surpassing traditional machine learning techniques in both segmentation and classification tasks. However, challenges regarding dataset size and diversity, model interpretability, and research focus need to be addressed. Future research should focus on creating larger, more balanced, and representative datasets, improving model explainability through visualization tools and techniques, expanding research to include a wider range of skin conditions, and exploring innovative model architectures such as Swin transformers and the incorporation of reinforcement learning and multimodal data.

The review's scope is limited to the literature available in English language databases. The findings are based on reported results from various studies, and the heterogeneity of datasets, methodologies, and evaluation metrics across these studies makes direct comparisons challenging. The absence of a standardized dataset and consistent evaluation protocols makes generalization of findings difficult. Finally, the review primarily focuses on the technical aspects of AI applications in dermatology, with less emphasis on the clinical implications and ethical considerations.