A strategy-based taxonomy of AI–radiomics integration in breast cancer diagnostics: A focused systematic review using a quantitative high-impact scoring framework

Abstract

Integrating artificial intelligence (AI) and radiomics is transforming breast cancer diagnostics; however, a clearly defined framework for evaluating distinct strategic approaches remains lacking. This systematic review addresses this gap by analysing 43,391 studies, rigorously filtering them to identify 10 seminal, high-impact articles published between January 2021 and May 2024. We introduce a novel five-category taxonomy for the systematic classification of state-of-the-art research based on the fundamental strategy of AI-radiomics integration: (1) Advanced Radiomic Feature Engineering, (2) Deep Feature Learning via Transfer Learning, (3) Hybrid Feature Fusion, (4) Multi-View/Multi-Modal Architectural Fusion, and (5) AI-in-Workflow & Clinical Integration. Our analysis reveals a clear evolutionary trend from feature-level optimisation towards deeply integrated systems that are architecturally complex and clinically aware. Key findings demonstrate that hybrid and architectural fusion models achieve the highest levels of diagnostic accuracy, with AUC values reaching 0.947 for malignancy classification (LA-Net, ultrasound cohort) and 0.979 for density estimation on the VinDr-Mammo dataset within architectural fusion models, while workflow-integration studies demonstrate quantifiable real-world impact; notably, a retrospective multicentre simulation involving 4209 women with BI-RADS 4 mammographic lesions demonstrated that a deep learning decision-support tool could potentially reduce unnecessary benign biopsies by up to 50% in low-to-moderate risk subgroups while maintaining near-perfect sensitivity (99.8%) for malignancy—a finding that requires prospective validation before broader clinical generalisation. We conclude that the future of the field lies in the synergistic application of these diverse strategies to create tools that are not only accurate but also interpretable, efficient, and trusted in clinical practice.