Please login or create an account. If you do not have access to this content, you will be shown a 30 second preview and licensing options.

Presentation

Dermatopathology, Melanoma Overdiagnosis, and the Role of AI in Refining Diagnosis

Description

Atam Oshiri argues that melanoma is overdiagnosed largely because dermatopathology has historically prioritized sensitivity over specificity, especially for early lesions. He reviews foundational work from Breslow, Ackerman, and the NIH consensus that established histologic criteria for melanoma and thin melanoma prognosis, but also created a broad diagnostic net that includes many very low-risk lesions. He highlights the poor reproducibility of borderline melanocytic diagnoses, noting substantial disagreement even among experts and between self-readings over time, which contributes to diagnostic drift and an ever-lower threshold for calling lesions melanoma or melanoma in situ. He suggests practical steps to reduce overdiagnosis: stop grouping melanoma in situ with invasive melanoma, curb diagnostic drift, and remove very low-risk lesions from the melanoma category, potentially renaming them as low malignant potential neoplasms. He then uses chess as an analogy for AI progress, contrasting human rules with AI systems like AlphaZero that learned novel strategies and surpassed conventional engines. In dermatopathology, he proposes that AI and molecular tools could reveal better predictive features, reduce noise, and help identify subsets of lesions now labeled melanoma that are actually low-risk, while histopathology remains the current gold standard.

View more

Conclusions

Melanoma is likely overdiagnosed in part because histopathology has prioritized sensitivity over specificity, especially for very early lesions.
Melanoma in situ should probably be separated from invasive melanoma because its clinical risk is very low and its inclusion inflates melanoma incidence.
Diagnostic drift over time appears to push borderline melanocytic lesions toward a melanoma label rather than away from it.
There is substantial interobserver and intraobserver variability in the classification of melanocytic lesions, including thin melanomas and dysplastic nevi.
Dermatopathologists, even when experienced, tend to diagnose higher-grade atypia or early melanoma more often than general pathologists.
A subset of very thin pT1a melanomas behaves so indolently that they may warrant a different, less alarming terminology.
Renaming very low-risk lesions as low-malignant-potential melanocytic neoplasms could reduce the harms of overdiagnosis.
Artificial intelligence and molecular tools may help refine melanoma diagnosis by identifying features and subgroups that histology alone currently overcalls.
AI may uncover prognostically important histologic features not emphasized by current dogma, such as nuclear size or other subtle patterns.
The future of melanoma diagnosis will likely require revising longstanding categorical assumptions rather than simply training pathologists harder.

Breslow, Ann Surg, 1970
Goldsmith et al., JAMA, 1992
Xavier-Junior and Ocanha-Xavier, Ann Diag Pathol, 2022
Elmore et al, BMJ, 2017
Barnhill et al, JAMA Path Lab Med, 2023
Welch et al, NEJM, 2021
Kerr et al, JAMA Derm, 2023
Elder et al, Clin Dermatol, 2024
De Logu et al, Front Oncol, 2020
Kim et al, JID, 2022