Can Artificial Intelligence Revolutionize Skin Cancer Detection in Clinical Settings?

🔍 Key Finding AI-powered tools for skin cancer detection and classification show promising results in terms of accuracy, sensitivity, and specificity, particularly for melanoma. However, broader clinical validation, standardization, and collaboration with dermatologists are crucial for effective and ethical implementation in real-world clinical settings.

🔬 Methodology Overview

• Design: Systematic review
• Data Sources: PubMed, Scopus, Embase, and Web of Science databases (up to April 4, 2023)
• Search Terms: (“skin cancer” OR “skin lesion” OR “dermatology” OR “dermatoscopy” OR “melanoma”) AND (“artificial intelligence” OR “neural network” OR “deep learning” OR “convolutional neural network” OR “transfer learning” OR “machine learning” OR “Computer aided diagnostic*” OR “CAD” OR “image classification” OR “image processing” OR “Internet of things” OR “Data mining” OR “Iot”) AND (“real-time” or “real time” OR “real-world” OR “real world” OR “smartphone”) AND NOT (“Meta-Analysis” OR “Meta Analysis” OR “Systematic Review”)
• Selection Criteria: Studies focusing on AI for skin cancer detection, classification, and assessment in clinical settings; journal articles and conference proceedings only.
• Data Extraction: Publication details, data characteristics (type, origin, quantity), AI techniques and resources, study findings, diversity considerations, accessibility information, and medical professional involvement.
• Analysis Approach: Narrative synthesis of extracted data.
• Scope: AI applications for skin cancer in real-world clinical environments.

📊 Results

• Data Diversity: Studies utilized various data types, including clinical (macroscopic) and dermoscopic images, sourced from databases (PH2, ISIC), directly from patients via mobile devices, and teledermatology projects. Dataset sizes varied significantly, from as few as 83 images to over 130,000. Some studies lacked clear data origin details, potentially impacting generalizability.

• Techniques and Classifiers: Image processing techniques included segmentation, feature extraction, and classification. Classifiers ranged from traditional machine learning algorithms (SVM, K-NN) to deep learning models (CNNs, YOLOv2, DenseNet, NASNetMobile). Some studies employed ensembles of multiple models.

• Performance Metrics: Studies reported varying performance metrics. Sensitivity for melanoma detection ranged from 80.5% to 100%, while specificity ranged from 45.5% to 100%. Overall accuracy varied from 75.07% to 97.5%. The YOLOv2 model achieved 86% overall accuracy, 86.35% recall, and 85.9% specificity for melanoma detection in dermoscopic images.

• Clinical Validation: Many studies lacked robust clinical validation, with some still in the testing phase. Direct comparison with dermatologist diagnoses was performed in some cases, revealing comparable or even superior performance by AI systems.

• Diversity and Collaboration: Some studies addressed ethnic diversity limitations by using locally generated data or stratifying by Fitzpatrick skin type. Collaboration with dermatologists was emphasized for data labeling, quality control, and validation, but the level of involvement varied.

• Accessibility and Availability: Mobile device implementation was a common goal, aiming to increase access to diagnostic support, particularly in underserved areas. However, practical accessibility and availability remained a challenge due to resource limitations and technical complexity.

💡 Clinical Impact AI-powered tools hold promise for revolutionizing early skin cancer detection and diagnosis, particularly in underserved areas with limited access to specialists. However, rigorous clinical validation and collaboration with dermatologists are crucial to ensure their safe and effective integration into real-world practice, balancing the potential benefits with the need for ongoing research and refinement.

🤔 Limitations

• Performance variation between devices.
• Possibility of unnecessary excisions.
• Most solutions have not been validated in clinical settings.
• Most solutions were not developed in collaboration with dermatologists and other healthcare professionals.
• Limited resource availability, technical complexity, or absence of clear guidelines for many systems.
• Need for broader collaboration among companies, accessibility experts, programmers, and users to improve accessibility.
• Ethical considerations regarding patient data privacy and security, and transparency in algorithm development.

✨ What It Means For You AI-powered tools hold promise for revolutionizing skin cancer detection by offering rapid, accurate analyses and increased accessibility to diagnostic support, particularly in underserved areas. However, rigorous clinical validation and collaboration with dermatologists are crucial to ensure these tools are reliable, effective, and ethically implemented in real-world practice.

Reference Furriel BCRS, Oliveira BD, Prôa R, Paiva JQ, Loureiro RM, Calixto WP, Reis MRC, Giavina-Bianchi M. Artificial intelligence for skin cancer detection and classification for clinical environment: a systematic review. Front. Med. 2024;10:1305954. https://doi.org/10.3389/fmed.2023.1305954