The Bench-to-Bedside Journey of Deep Learning in Dermato-Oncology: A Systematic Review of Clinical Readiness Across the Translational Continuum

Deep learning systems have matched dermatologists on curated skin cancer image sets for almost a decade, but evidence from patient care has accumulated slowly and unevenly. This systematic review mapped where that evidence now sits along the translational continuum. A reproducible PubMed search run in August 2026 identified 656 records; after single-screener selection against pre-specified criteria and full-text assessment of 39 reports, 29 clinical-stage studies (prospective, controlled, or real-world evaluations of AI for cutaneous malignancy in patients) were included and staged as prospective standalone accuracy (n = 19), prospective clinician-with-AI use (n = 5), controlled evaluation (n = 1), or real-world deployment (n = 4). Risk of bias, appraised with QUADAS-2, ROBINS-I, or JBI tools, was high or serious in 24 studies, most often because of partial verification and preselected lesion spectra, and the certainty of every principal finding was low or very low (GRADE). No included study was a randomised trial with patient outcomes, and none reported stage at diagnosis, mortality, or overdiagnosis. Across modalities, sensitivity held up in prospective use, commonly above 90% at the pre-set operating point, while specificity fell to between 9% and 85% and was frequently lower than that of the clinicians it was compared against. Performance moved with acquisition hardware, lesion spectrum, and skin type, and one market-approved dermoscopic classifier reached areas under the curve of 0.54 to 0.62 in two Australian primary care studies. One product family progressed from prospective accuracy to national-pathway deployment under a conditional health technology recommendation, with human second reading overturning a large share of automated discharges. We argue that the reported specificity loss reflects safety-first operating points, spectrum preselection, and the choice of comparator rather than model failure alone, and propose readiness criteria linking each translational stage to the evidence needed before autonomous or primary-care use. The review followed PRISMA 2020, was not registered, and received no external funding.