MRI radiomics for predicting pathological complete response in breast cancer: a systematic review of current evidence and limitations.
Authors
Affiliations (6)
Affiliations (6)
- Barcelona Artificial Intelligence in Medicine Lab (BCN-AIM), Facultat de Matemàtiques i Informàtica, Universitat de Barcelona, Barcelona, Spain; Department of Radiology, University of Cambridge, Cambridge, United Kingdom.
- 2nd Department of Radiology, Medical University of Gdansk, Gdansk, Poland. Electronic address: [email protected].
- 2nd Department of Radiology, Medical University of Gdansk, Gdansk, Poland.
- Barcelona Artificial Intelligence in Medicine Lab (BCN-AIM), Facultat de Matemàtiques i Informàtica, Universitat de Barcelona, Barcelona, Spain.
- Department of Electrical and Computer Engineering, Hellenic Mediterranean University, Heraklion, Greece; Institute of Computer Science, Foundation for Research and Technology Hellas (FORTH), Heraklion, Greece.
- Barcelona Artificial Intelligence in Medicine Lab (BCN-AIM), Facultat de Matemàtiques i Informàtica, Universitat de Barcelona, Barcelona, Spain; Institució Catalana de Recerca i Estudis Avançats (ICREA), Barcelona, Spain.
Abstract
Machine learning models based on MRI radiomics show promise for predicting pathological complete response (pCR) to neoadjuvant therapy in breast cancer, with increasing interest in clinical translation. However, uncertainties remain regarding methodological quality, generalizability, and performance across patient subgroups. This systematic review provides a comprehensive evaluation of MRI radiomics models for pCR prediction, focusing on methodological robustness, validation practices, and factors affecting equitable performance. The review protocol was registered with PROSPERO (Registration ID: CRD420251143053). A systematic search of five electronic databases identified studies evaluating MRI-based radiomics models for pCR prediction in breast cancer. Methodological quality was assessed using the METRICS score. Fairness and risk of bias were evaluated using a custom framework adapted from QUADAS-2, focusing on demographic reporting, sample size adequacy, validation strategy, and subgroup performance analysis. Thirty-five studies met the inclusion criteria (sample sizes 55-442). Most were retrospective, single-center, and methodologically heterogeneous. Although reported discrimination was favorable (pooled AUC 0.813, moderate heterogeneity [I<sup>2</sup> = 46.3%, τ<sup>2</sup> = 0.081]), 86% of studies were statistically underpowered and external validation was uncommon. Only 14% evaluated performance across clinically relevant subgroups. Reporting of key demographic and clinical variables was inconsistent, limiting assessment of generalizability and equitable model performance. MRI radiomics models for pCR prediction demonstrate technical potential but face important methodological and translational limitations. Future studies should prioritize prospective multi-center designs, adequate sample sizes, transparent reporting, and subgroup analyses. Improved demographic reporting and subgroup-specific evaluation are needed to enable robust assessment of equitable model performance before clinical deployment.