Artificial Intelligence for Borderline Breast Lesions on Whole-Slide Images: A Systematic Review of Accuracy, ncertainty and Generalisation

Abstract Background : Borderline breast lesions, such as atypical ductal hyperplasia (ADH), flat epithelial atypia (FEA) and lobular neoplasia, are among the least reproducible diagnoses in breast pathology. We reviewed how well artificial intelligence (AI) on digitised haematoxylin and eosin slides classifies these lesions, whether it was tested at other hospitals, and whether it reports uncertainty. Methods : Following PRISMA 2020 and a registered protocol, we searched PubMed, OpenAlex and Europe PMC to 22 October 2026, with citation and preprint searching. Screening was assisted by a large language model and checked by both authors. Risk of bias was assessed with QUADAS-2, and results were synthesised narratively (SWiM). Results : Eighteen reports (2016 to 2026) from only seven patient cohorts were included. In multi-class settings, AI performed poorly on ADH (recall 0.19 to 0.69; F1 0.24 to 0.47 on the public BRACS dataset, where pathologists reached 0.38). In binary tasks, AI separated ductal carcinoma in situ (DCIS) from atypia with accuracy 0.85 to 0.90, versus 0.80 for pathologists on the same cases. Only 2 of 18 reports tested the model at another hospital (external area under the curve 0.92 and 0.61 to 0.70). No report offered calibrated uncertainty or referral of uncertain cases to a pathologist. Risk of bias from patient selection was high in all reports. Conclusions : AI for borderline breast lesions is promising in narrow tasks but weak for ADH in realistic multi-class settings, rarely validated externally, and never designed to say when it is unsure. Future studies need multi-centre test sets and calibrated uncertainty with a refer-to-pathologist option.