Can Machine Learning Revolutionize Skin Lesion Diagnosis and Treatment?
🔍 Key Finding This review summarizes machine learning techniques for skin lesion classification, segmentation, and detection, finding that deep learning models, particularly when combined with preprocessing techniques or other classifiers, offer high accuracy on benchmark datasets like HAM10000 and ISIC, showing promise for aiding dermatological diagnosis. However, challenges remain regarding generalizability, interpretability, and the need for larger, more diverse datasets.
🔬 Methodology Overview
- Design: Systematic review of published research articles.
- Research Questions: Five pre-defined research questions focusing on machine learning in skin disease diagnosis (targets, techniques, datasets, success rates, and future challenges).
- Search Strings: Defined search strings combining keywords related to skin conditions, machine learning methods, and diagnostic targets (detection, classification, segmentation).
- Inclusion Criteria: Papers focused on skin disease/cancer, including skin cancer/melanoma; published in reputable journals (impact factor, indexed in Web of Science/Scopus/conference proceedings); written in English; published between 2020-2023 (except Sections 1 and 4, which had no publication year restriction).
- Exclusion Criteria: Non-peer-reviewed publications, abstracts, editorials, book reviews, scientific reports; M.Sc./Ph.D. theses, posters, seminars; case studies; publications before 2020 (except Sections 1 and 4).
- Data Sources: IEEE Xplore, MDPI, Google Scholar, Springer Link, and ScienceDirect databases.
- Analysis Approach: Qualitative synthesis of findings from selected studies, summarizing methodologies, datasets, performance metrics, contributions, and limitations.
📊 Results
- EfficientNets for Skin Cancer Classification: Achieved 87.9% accuracy, 88% precision, 88% recall, 87% F1-score, and 97.53% AUC on the HAM10000 dataset.
- Deep Learning with MobileNet V2 and LSTM: Achieved 90.21% accuracy on the HAM10000 dataset, outperforming VGG16, AlexNet, and other traditional machine learning methods.
- CNN and Machine Learning for Skin Lesion Classification: Achieved 95.18% accuracy on the HAM10000 dataset, outperforming several state-of-the-art systems.
- DenseNet and ConvNeXt Fusion: Achieved 95.29% accuracy on HAM10000 and 96.54% accuracy on a dataset from Peking Union Medical College Hospital.
- Hybrid Deep Learning Models and Machine Learning Classifiers: Achieved 99.94% accuracy on the HAM10000 dataset using DenseNet201 combined with Logistic Regression.
- Optimized Region Growing and Autoencoder Classification: Achieved 94.2% accuracy on the PH2 dataset.
- Deep Bottleneck Transformer Model: Achieved 92.1% accuracy, 90.1% sensitivity, and 91.9% specificity on ISIC2017 and 95.84% accuracy and 96.1% precision on HAM10000.
💡 Clinical Impact Automated skin disease classification and detection using machine learning, particularly deep learning, shows promise for improving diagnostic accuracy and speed compared to traditional visual inspection, potentially enabling earlier diagnosis and treatment of skin cancers and other skin conditions. This could lead to wider adoption of AI-powered diagnostic tools as a complement to dermatologist assessments, improving patient outcomes and access to care.
🤔 Limitations
- Reliance on dermoscopic images limits applicability to clinical images.
- Many models lack interpretability, hindering clinical trust and understanding.
- Generalizability to diverse datasets and real-world scenarios needs further validation.
- Large labeled datasets are required for training, which may not always be available.
- Potential biases in training data may affect diagnostic accuracy.
- Evaluation primarily focuses on accuracy and may not fully capture clinical relevance.
- Integration of AI systems into existing healthcare workflows requires further research.
✨ What It Means For You This review highlights the rapid advancement of machine learning techniques for analyzing skin lesions, offering doctors a potential tool to improve diagnostic accuracy and efficiency in detecting and classifying skin diseases, including melanoma. Automated analysis of dermoscopic and macroscopic images can complement clinical evaluations, potentially reducing reliance on invasive biopsies and expediting treatment for patients. Further research is needed to validate these techniques in clinical settings and address challenges like data bias and interpretability before widespread adoption.
Reference Debelee TG. Skin Lesion Classification and Detection Using Machine Learning Techniques: A Systematic Review. Diagnostics. 2023;13:3147. https://doi.org/10.3390/diagnostics13193147