Can Artificial Intelligence Revolutionize Non-Melanoma Skin Cancer Care?

by Haroon Ahmad, MD 2025-01-01 00:00
PhysicianPractice Innovation

🔍 Key Finding AI demonstrates potential for diagnosing and treating non-melanoma skin cancers, but its real-world performance lags behind theoretical capabilities due to limitations in training data, particularly the lack of diverse skin tones, and the need for standardized evaluation guidelines. Further development of AI for skin cancer requires high-quality, diverse datasets curated by dermatologists and transparent reporting of algorithm development and performance.

🔬 Methodology Overview

  • Design: Narrative review/Opinion statement
  • Data Sources: Published literature on artificial intelligence and non-melanoma skin cancer.
  • Selection Criteria: Focus on clinical applications of AI in NMSC diagnosis and treatment, including BCC, SCC, and MCC. Also addresses challenges and future directions, such as guideline development and addressing skin of color disparities.
  • Analysis Approach: Qualitative synthesis of findings and expert opinion on current state and potential of AI in NMSC.
  • Scope: Clinical application of AI in dermatology, specifically for non-melanoma skin cancers. Includes discussion of diagnostic tools, treatment design, and data curation needs.

📊 Results

  • AI using dermatoscope images achieved 95.3% accuracy in diagnosing NMSC, compared to 75.3% accuracy with smartphone images in a real-world telehealth setting. Specificity was similar for both.
  • Supervised machine learning algorithms achieved 45.1% accuracy in predicting multidisciplinary team treatment choices (conventional, surgery, or radiotherapy) for complex BCC and 37.5% accuracy in predicting Mohs surgery referral.
  • An algorithm predicting response to radiotherapy for SCC achieved 85.7% sensitivity, 97.6% specificity, and 91.7% overall accuracy.
  • AI trained on artificially darkened images of light-skinned patients showed improved sensitivity, specificity, positive predictive value, and negative predictive value when diagnosing skin cancer in darker skin tones.
  • A study in a Hawaiian multiethnic population found that AI performance was negatively impacted by the lack of diversity in skin types (Fitzpatrick skin types 4-6) in training datasets.
  • “CLEAR criteria” guidelines propose 25 items for reporting on image-based AI development and assessment in dermatology to improve transparency and evaluation.

💡 Clinical Impact AI may improve the time efficiency of skin cancer diagnosis, particularly for non-melanoma skin cancers, and may facilitate care in settings with limited dermatology expertise. Further development focusing on diverse training datasets and transparent algorithms is needed to optimize real-world performance and address existing disparities in diagnosis.

🤔 Limitations

  • Simple errors by the AI, such as a diagnostic decision inappropriately influenced by surgical skin markings or other artifacts found in training datasets.
  • Possible misdiagnoses in the clinic.
  • Increased false positives or negatives in AI performance studies.
  • Decreased validity.
  • Limited incidence resulting in fewer images for algorithm training and worse AI performance for NMSC detection among darker skin tones.
  • Individuals and clinical scenarios that fall outside of the characteristics included in training datasets may not be reliably diagnosed by AI.

✨ What It Means For You AI can improve the diagnosis and treatment of non-melanoma skin cancers, potentially increasing efficiency and extending care to underserved areas. However, dermatologists must actively curate diverse and high-quality datasets to train these AI algorithms and ensure accurate and equitable performance across all skin tones. Furthermore, transparent guidelines and validation are crucial for building trust and facilitating wider adoption in clinical practice.

Reference Sanchez K, Kamal K, Manjaly P, Ly S, Mostaghimi A. Clinical Application of Artificial Intelligence for Non-melanoma Skin Cancer. Curr Treat Options Oncol. 2023;24:373–379. https://doi.org/10.1007/s11864-023-01065-4