264P When AI meets the probe: A systematic review of diagnostic accuracy in handheld breast ultrasonography

Abstract

Background: Early melanoma detection is essential for patient survival, yet diagnostic accuracy varies across settings.Artificial intelligence (AI) systems, including convolutional neural networks (CNNs), are increasingly evaluated in prospective trials as adjuncts or alternatives to clinician assessment.This systematic review assessed trial evidence for AI in suspected melanoma, distinguishing between AIalone (AI versus human) and AI-assisted (AI plus human versus human) approaches.Methods: PubMed, Embase, and the Cochrane Central Register of Controlled Trials were searched.A total of 33 deduplicated records were obtained from these databases, and 6 were included for full text review.Population (P) comprised adults with suspected melanoma.Intervention (I) was AI use, either AI-alone or AI-assisted.Comparator (C) was standard clinician care.Outcomes (O) included diagnostic performance metrics (sensitivity, specificity, area under the receiver operating characteristic curve [AUROC]) and concordance.A meta-analysis in RStudio pooled diagnostic outcomes, including AI-alone, AI-assisted, and human-only arms.Results: 3 prospective clinical trials were included in our final analysis.Investigated AI systems ranged from feed-forward to convolutional neural networks.Meta-analysis of 2 trials showed a pooled sensitivity of 75.5% AI melanoma detection, and 69.8% for human readers, compared to 86.5% for AI-assisted detection.In contrast, for AI alone the pooled specificity was higher, at 90.0% when compared to AIassisted detection.Limitations included small melanoma numbers, exclusion of darker skin types, and restricted generalisability.Conclusions: Clinical trial evidence indicates that AI diagnostic tools achieve accuracy comparable to expert dermatologists and enhance performance of non-expert clinicians in detecting melanoma.Although promising, larger multicentre trials with greater ethnic and lesion diversity are required to confirm reproducibility, assess patient outcomes, and define regulatory pathways for clinical implementation.

Ähnliche Arbeiten