Breast Tumor Segmentation in Automated Breast Ultrasound: A Scoping Review of Deep Learning Approaches (2020–2026)

Automated breast ultrasound (ABUS) acquires whole-breast volumes that can be reviewed independently of acquisition, but tumor segmentation in these volumes remains difficult. This scoping review maps deep learning methods for tumor segmentation in ABUS and examines how they were evaluated. Following PRISMA-ScR, PubMed, IEEE Xplore and ScienceDirect were searched for studies published between January 2020 and October 2026, without an open access restriction. Of 65 records identified, 32 sources were included: 27 primary segmentation studies, four segmentation-integrated computer-aided diagnosis studies, and the report of the TDSC-ABUS 2023 challenge, which describes 14 team methods evaluated on a common hidden test set. The primary studies were grouped into six families: CNN-based, attention-based, hybrid CNN–Transformer, lightweight/architecture-search/distillation, semi- and weakly supervised, and foundation-model-based approaches. Twelve studies used the same single-center dataset, one used more than one scanner vendor, and four tested on an external center. Reported Dice values depended strongly on the evaluation setting: values of 0.84 or higher were obtained on 2D slices, on localized regions (including 3D lesion-containing crops) or on small cohorts, whereas whole-volume evaluations on public data gave 0.58–0.65 and challenge entries 0.31–0.61. Self-configuring CNN pipelines gave the best challenge results, and most foundation-model approaches required adaptation and prompts. No family has been shown to outperform the others. Progress depends on shared multi-center benchmarks, standardized reporting of evaluation settings, data-efficient and privacy-preserving training, and evaluation of runtime and clinical workflow.