Trend AnalysisMedicine & Health
AI in Dermatology: Deep Learning for Skin Cancer Screening
Melanoma, the deadliest skin cancer, has a ~100% five-year survival rate when detected at stage I but approximately 35% at stage IV (current SEER data, reflecting immunotherapy-era improvements). Earl...
By Sean K.S. Shin
This blog summarizes research trends based on published paper abstracts. Specific numbers or findings may contain inaccuracies. For scholarly rigor, always consult the original papers cited in each post.
The Question
Melanoma, the deadliest skin cancer, has a ~100% five-year survival rate when detected at stage I but approximately 35% at stage IV (current SEER data, reflecting immunotherapy-era improvements). Early detection is life-saving, yet dermatologists are scarce (many countries have fewer than 1 per 100,000 population). AI models have demonstrated dermatologist-level accuracy in classifying skin lesions from dermoscopic and clinical photographs. Can AI-powered screening tools โ on smartphones, in primary care clinics, or integrated into telemedicine โ democratise expert-level skin cancer detection?
Landscape
Kalidindi (2024) reviewed AI's role in melanoma diagnosis, documenting that CNN-based models reportedly achieve high sensitivity and specificity on benchmark datasets (ISIC, HAM10000). However, performance drops significantly on images from different populations, equipment, and clinical settings โ the classic domain shift problem in medical AI.
Ornek et al. (2024) used the ISIC 2019 dataset (8 skin lesion categories) with SqueezeNet-based feature extraction followed by multiple classifiers (ANN, kNN, Random Forest, Logistic Regression), achieving up to 71.8% classification accuracy (ANN) across multiple lesion types. Semerci et al. (2024) extended AI applications to molecular classification of head and neck skin cancers, integrating AI-based histopathological analysis with genomic data.
Trager et al. (2024) reviewed AI for non-melanoma skin cancers (basal cell carcinoma, squamous cell carcinoma) โ among the top five most common cancers globally but often overlooked in AI dermatology research, which disproportionately focuses on melanoma. Their analysis identified AI-assisted surgical margin assessment as a promising clinical application beyond screening.
Ruga et al. (2025) developed MultiExCam, an explainable AI architecture that combines multiple attention mechanisms with saliency mapping to provide clinically interpretable explanations for AI classifications โ addressing the "black box" barrier to clinical adoption.
Key Claims & Evidence
<
| Claim | Evidence | Verdict |
|---|
| AI achieves dermatologist-level melanoma detection on benchmark datasets | High sensitivity and specificity reported across studies (Kalidindi 2024) | Supported on curated data; real-world performance lower |
| Domain shift degrades AI performance across populations and equipment | Significant accuracy drops on out-of-distribution images (Kalidindi 2024) | Confirmed; the critical barrier to clinical deployment |
| AI is applicable beyond melanoma to NMSC and surgical margin assessment | BCC and SCC classification and Mohs surgery assistance (Trager et al. 2024) | Emerging; less mature than melanoma AI |
| Explainable AI improves clinical trust | Multi-attention saliency maps highlight diagnostically relevant features (Ruga et al. 2025) | Demonstrated; clinical impact on decision-making not yet measured |
Open Questions
Skin colour equity: Most training datasets are dominated by light-skinned individuals. Can AI models perform equitably across Fitzpatrick skin types IโVI?
Smartphone-based screening: Can consumer smartphone cameras achieve the image quality needed for reliable AI classification, or is dermoscopy equipment required?
Overdiagnosis risk: If AI screening increases biopsy rates, will the ratio of unnecessary biopsies to detected cancers be acceptable?
Integration into clinical workflow: Should AI serve as a triage tool (flagging high-risk lesions for dermatologist review) or a diagnostic tool (providing classification directly to patients)?Referenced Papers
- [1] Kalidindi, S. (2024). The Role of AI in the Diagnosis of Melanoma. Cureus, 16(9), e69818. DOI: 10.7759/cureus.69818
- [2] Ornek, H.K. et al. (2024). Deep Learning-Based Skin Lesion Classification for Melanoma. DOI: 10.58190/imiens.2024.101
- [3] Semerci, Z.M. et al. (2024). AI in Early Diagnosis of Head and Neck Skin Cancers. Diagnostics, 14(14), 1477. DOI: 10.3390/diagnostics14141477
- [4] Trager, M. et al. (2024). AI for Non-Melanoma Skin Cancer. Clinics in Dermatology. DOI: 10.1016/j.clindermatol.2024.06.016
- [5] Ruga, T. et al. (2025). MultiExCam: Explainable AI for Skin Lesion Classification. Computer Methods and Programs in Biomedicine. DOI: 10.1016/j.cmpb.2025.109081
๋ฉด์ฑ
์กฐํญ: ์ด ๊ฒ์๋ฌผ์ ์ ๋ณด ์ ๊ณต ๋ชฉ์ ์ ์ฐ๊ตฌ ๋ํฅ ๊ฐ์์ด๋ค. ํ์ ์ ์๋ฌผ์์ ์ธ์ฉํ๊ธฐ ์ ์ ๊ตฌ์ฒด์ ์ธ ์ฐ๊ตฌ ๊ฒฐ๊ณผ, ํต๊ณ, ์ฃผ์ฅ์ ์๋ฌธ ๋
ผ๋ฌธ๊ณผ ๋์กฐํ์ฌ ๊ฒ์ฆํด์ผ ํ๋ค.
AI์ ํผ๋ถ๊ณผํ: ํผ๋ถ์ ์ ๋ณ ๊ฒ์ฌ๋ฅผ ์ํ ๋ฅ๋ฌ๋
๋ถ์ผ: ์ํ | ๋ฐฉ๋ฒ๋ก : ์ ์ฐ-์์
์ ์: Sean K.S. Shin | ๋ ์ง: 2026-03-17
์ฐ๊ตฌ ์ง๋ฌธ
๊ฐ์ฅ ์น๋ช
์ ์ธ ํผ๋ถ์์ธ ํ์์ข
(melanoma)์ 1๊ธฐ์ ๋ฐ๊ฒฌ๋ ๊ฒฝ์ฐ 5๋
์์กด์จ์ด ~100%์ ๋ฌํ์ง๋ง, 4๊ธฐ์์๋ ์ฝ 35%์ ๋ถ๊ณผํ๋ค(๋ฉด์ญ์๋ฒ ์๋์ ๊ฐ์ ์ ๋ฐ์ํ ํํ SEER ๋ฐ์ดํฐ). ์กฐ๊ธฐ ๋ฐ๊ฒฌ์ ์๋ช
์ ๊ตฌํ์ง๋ง, ํผ๋ถ๊ณผ ์ ๋ฌธ์๋ ๋ถ์กฑํ ์ค์ ์ด๋ค(๋ง์ ๊ตญ๊ฐ์์ ์ธ๊ตฌ 10๋ง ๋ช
๋น 1๋ช
๋ฏธ๋ง). AI ๋ชจ๋ธ์ ํผ๋ถ๊ฒฝ(dermoscopic) ์ฌ์ง ๋ฐ ์์ ์ฌ์ง์ ํตํ ํผ๋ถ ๋ณ๋ณ ๋ถ๋ฅ์์ ํผ๋ถ๊ณผ ์ ๋ฌธ์ ์์ค์ ์ ํ๋๋ฅผ ์
์ฆํ ๋ฐ ์๋ค. ์ค๋งํธํฐ, 1์ฐจ ์ง๋ฃ ํด๋ฆฌ๋, ๋๋ ์๊ฒฉ ์๋ฃ์ ํตํฉ๋ AI ๊ธฐ๋ฐ ์ ๋ณ ๊ฒ์ฌ ๋๊ตฌ๊ฐ ์ ๋ฌธ๊ฐ ์์ค์ ํผ๋ถ์ ํ์ง๋ฅผ ๋์คํํ ์ ์์๊น?
์ฐ๊ตฌ ํํฉ
Kalidindi(2024)๋ ํ์์ข
์ง๋จ์์ AI์ ์ญํ ์ ๊ฒํ ํ๋ฉด์, CNN ๊ธฐ๋ฐ ๋ชจ๋ธ์ด ๋ฒค์น๋งํฌ ๋ฐ์ดํฐ์
(ISIC, HAM10000)์์ ๋์ ๋ฏผ๊ฐ๋(sensitivity)์ ํน์ด๋(specificity)๋ฅผ ๋ฌ์ฑํ๋ค๊ณ ๋ณด๊ณ ํ์๋ค. ๊ทธ๋ฌ๋ ์์ดํ ์ธ๊ตฌ ์ง๋จ, ์ฅ๋น, ์์ ํ๊ฒฝ์์๋ ์ฑ๋ฅ์ด ํฌ๊ฒ ์ ํ๋๋๋ฐ, ์ด๋ ์๋ฃ AI์์ ๋ํ๋๋ ์ ํ์ ์ธ ๋๋ฉ์ธ ์ด๋(domain shift) ๋ฌธ์ ์ด๋ค.
Ornek et al.(2024)์ ISIC 2019 ๋ฐ์ดํฐ์
(ํผ๋ถ ๋ณ๋ณ 8๊ฐ ๋ฒ์ฃผ)์ SqueezeNet ๊ธฐ๋ฐ ํน์ง ์ถ์ถ๊ณผ ๋ค์์ ๋ถ๋ฅ๊ธฐ(ANN, kNN, Random Forest, Logistic Regression)๋ฅผ ๊ฒฐํฉํ์ฌ ๋ค์ค ๋ณ๋ณ ์ ํ์์ ์ต๋ 71.8%์ ๋ถ๋ฅ ์ ํ๋(ANN ๊ธฐ์ค)๋ฅผ ๋ฌ์ฑํ์๋ค. Semerci et al.(2024)์ AI ๊ธฐ๋ฐ ์กฐ์ง๋ณ๋ฆฌํ์ ๋ถ์์ ์ ์ ์ฒด ๋ฐ์ดํฐ์ ํตํฉํ์ฌ ๋๊ฒฝ๋ถ ํผ๋ถ์์ ๋ถ์์ ๋ถ๋ฅ์ AI ์ ์ฉ ๋ฒ์๋ฅผ ํ์ฅํ์๋ค.
Trager et al.(2024)์ ๋นํ์์ข
ํผ๋ถ์(basal cell carcinoma, squamous cell carcinoma)์ ๋ํ AI ์ฐ๊ตฌ๋ฅผ ๊ฒํ ํ์๋ค. ์ด ๋ ์์ข
์ ์ ์ธ๊ณ์ ์ผ๋ก ๊ฐ์ฅ ํํ 5๋ ์์ ์ํ์ง๋ง, ํ์์ข
์ ๋ถ๊ท ํ์ ์ผ๋ก ์ง์ค๋ AI ํผ๋ถ๊ณผํ ์ฐ๊ตฌ์์๋ ๊ฐ๊ณผ๋๋ ๊ฒฝ์ฐ๊ฐ ๋ง๋ค. ๊ทธ๋ค์ ๋ถ์์ AI ๋ณด์กฐ ์์ ์ ์ ์ฐ(surgical margin) ํ๊ฐ๋ฅผ ์ ๋ณ ๊ฒ์ฌ๋ฅผ ๋์ด์ ์ ๋งํ ์์ ์ ์ฉ ๋ถ์ผ๋ก ํ์ธํ์๋ค.
Ruga et al.(2025)์ ๋ค์ค ์ดํ
์
๋ฉ์ปค๋์ฆ(attention mechanism)๊ณผ ํ์ ์ฑ ๋งคํ(saliency mapping)์ ๊ฒฐํฉํ์ฌ AI ๋ถ๋ฅ์ ๋ํ ์์์ ์ผ๋ก ํด์ ๊ฐ๋ฅํ ์ค๋ช
์ ์ ๊ณตํ๋ ์ค๋ช
๊ฐ๋ฅํ AI(explainable AI) ์ํคํ
์ฒ์ธ MultiExCam์ ๊ฐ๋ฐํ์์ผ๋ฉฐ, ์ด๋ฅผ ํตํด ์์ ๋์
์ ์ฅ๋ฒฝ์ธ "๋ธ๋๋ฐ์ค" ๋ฌธ์ ์ ๋์ํ์๋ค.
์ฃผ์ ์ฃผ์ฅ ๋ฐ ๊ทผ๊ฑฐ
<
| ์ฃผ์ฅ | ๊ทผ๊ฑฐ | ํ์ |
|---|
| AI๋ ๋ฒค์น๋งํฌ ๋ฐ์ดํฐ์
์์ ํผ๋ถ๊ณผ ์ ๋ฌธ์ ์์ค์ ํ์์ข
ํ์ง๋ฅผ ๋ฌ์ฑํ๋ค | ์ฌ๋ฌ ์ฐ๊ตฌ์์ ๋์ ๋ฏผ๊ฐ๋ ๋ฐ ํน์ด๋ ๋ณด๊ณ (Kalidindi 2024) | ์ ์ ๋ ๋ฐ์ดํฐ์์๋ ์ง์ง๋จ; ์ค์ ํ๊ฒฝ์์์ ์ฑ๋ฅ์ ๋ ๋ฎ์ |
| ๋๋ฉ์ธ ์ด๋์ผ๋ก ์ธํด ์ธ๊ตฌ ์ง๋จ ๋ฐ ์ฅ๋น ๊ฐ AI ์ฑ๋ฅ์ด ์ ํ๋๋ค | ๋ถํฌ ์ธ(out-of-distribution) ์ด๋ฏธ์ง์์ ์ ์๋ฏธํ ์ ํ๋ ์ ํ(Kalidindi 2024) | ํ์ธ๋จ; ์์ ๋ฐฐ์น์ ํต์ฌ ์ฅ๋ฒฝ |
| AI๋ ํ์์ข
์ ๋์ด ๋นํ์์ข
ํผ๋ถ์(NMSC) ๋ฐ ์์ ์ ์ ์ฐ ํ๊ฐ์๋ ์ ์ฉ ๊ฐ๋ฅํ๋ค | BCC ๋ฐ SCC ๋ถ๋ฅ์ ๋ชจ์ค ์์ (Mohs surgery) ๋ณด์กฐ(Trager et al. 2024) | ๋ฐ์ ์ค; ํ์์ข
AI๋ณด๋ค ์ฑ์๋๊ฐ ๋ฎ์ |
| ์ค๋ช
๊ฐ๋ฅํ AI๊ฐ ์์์ ์ ๋ขฐ๋ฅผ ํฅ์์ํจ๋ค | ๋ค์ค ์ดํ
์
ํ์ ์ฑ ๋งต์ด ์ง๋จ์ ์ผ๋ก ๊ด๋ จ๋ ํน์ง์ ๊ฐ์กฐ(Ruga et al. 2025) | ์
์ฆ๋จ; ์์ฌ ๊ฒฐ์ ์ ๋ํ ์์์ ์ํฅ์ ์์ง ์ธก์ ๋์ง ์์ |
๋ฏธํด๊ฒฐ ์ง๋ฌธ
ํผ๋ถ์ ํํ์ฑ: ๋๋ถ๋ถ์ ํ๋ จ ๋ฐ์ดํฐ์
์ ๋ฐ์ ํผ๋ถํค์ ๊ฐ์ง ๊ฐ์ธ๋ค๋ก ํธํฅ๋์ด ์๋ค. AI ๋ชจ๋ธ์ด Fitzpatrick ํผ๋ถ ์ ํ IโVI ์ ๋ฐ์ ๊ฑธ์ณ ํํ์ฑ ์๊ฒ ์ฑ๋ฅ์ ๋ฐํํ ์ ์์๊น?
์ค๋งํธํฐ ๊ธฐ๋ฐ ์ ๋ณ ๊ฒ์ฌ: ์๋น์์ฉ ์ค๋งํธํฐ ์นด๋ฉ๋ผ๊ฐ ์ ๋ขฐํ ์ ์๋ AI ๋ถ๋ฅ์ ํ์ํ ์ด๋ฏธ์ง ํ์ง์ ๋ฌ์ฑํ ์ ์์๊น, ์๋๋ฉด ํผ๋ถ๊ฒฝ ์ฅ๋น๊ฐ ํ์ํ ๊น?
๊ณผ์ ์ง๋จ ์ํ: AI ์ ๋ณ ๊ฒ์ฌ๊ฐ ์๊ฒ(biopsy) ๋น์จ์ ๋์ธ๋ค๋ฉด, ๋ถํ์ํ ์๊ฒ ๋ ํ์ง๋ ์์ ๋น์จ์ด ์์ฉ ๊ฐ๋ฅํ ์์ค์ผ๊น?
์์ ์ํฌํ๋ก์ฐ๋ก์ ํตํฉ: AI๋ ๋ถ๋ฅ(triage) ๋๊ตฌ(๊ณ ์ํ ๋ณ๋ณ์ ํผ๋ถ๊ณผ ์ ๋ฌธ์ ๊ฒํ ๋ฅผ ์ํด ํ์)๋ก ๊ธฐ๋ฅํด์ผ ํ๋๊ฐ, ์๋๋ฉด ์ง๋จ ๋๊ตฌ(ํ์์๊ฒ ์ง์ ๋ถ๋ฅ๋ฅผ ์ ๊ณต)๋ก ๊ธฐ๋ฅํด์ผ ํ๋๊ฐ?References (5)
Kalidindi, S. (2024). The Role of Artificial Intelligence in the Diagnosis of Melanoma. Cureus.
ORNEK, H. K., YILMAZ, B., YASIN, E., & KOKLU, M. (2024). Deep Learning-Based Classification of Skin Lesion Dermoscopic Images for Melanoma Diagnosis. Intelligent Methods in Engineering Sciences.
Semerci, Z. M., Toru, H. S., รobankent Aytekin, E., Tercanlฤฑ, H., Chiorean, D. M., Albayrak, Y., et al. (2024). The Role of Artificial Intelligence in Early Diagnosis and Molecular Classification of Head and Neck Skin Cancers: A Multidisciplinary Approach. Diagnostics, 14(14), 1477.
Trager, M. H., Gordon, E. R., Breneman, A., Weng, C., & Samie, F. H. (2024). Artificial intelligence for nonmelanoma skin cancer. Clinics in Dermatology, 42(5), 466-476.
Ruga, T., Caroprese, L., Vocaturo, E., & Zumpano, E. (2026). MultiExCam: A multi approach and explainable artificial intelligence architecture for skin lesion classification. Computer Methods and Programs in Biomedicine, 273, 109081.