This systematic review investigated the biological underpinnings of MRI-derived radiomics with deep learning biomarkers for prognostic forecasting in breast cancer and critically evaluated the methodological quality across existing studies. The databases of PubMed, Embase, Web of Science Core Collection, and Cochrane Library were queried for studies from inception to April 2024. This report focused on research investigating the biological links of prognostic imaging biomarkers in breast cancer. The quality of each study was independently assessed using the Radiomics Quality Score (RQS) and the Prediction Model Risk of Bias Assessment Tool (PROBAST). Fourteen retrospective cohort studies encompassing 9,884 patients and spanning from 2020 to 2024 were included, along with one prospective validation cohort. Manual segmentation was the most common segmentation method (57% of studies), and half of the studies utilized multicenter data. The primary biological information was composed of novel biomarkers such as genes and pathways (64% of analyses) and traditional biomarkers such as histopathological factors (36% of analyses). The median RQS was 16.9 (range 7–23), reflecting moderate methodological quality. All studies exhibited a high risk of overall bias according to PROBAST criteria. Studies have demonstrated that the information captured by radiomics-based breast cancer prognosis models is closely correlated with genomic and histopathological factors, focusing on lncRNA expression, signaling pathways, and PD-L1 levels. Future advancements in biological interpretation may enhance the predictive accuracy of these models, ultimately improving breast cancer prognosis. The primary limitations of this study include the predominance of retrospective studies (13 out of 14), the overwhelming presence of Chinese data (11 out of 14), and the substantial risk of bias across the included studies. The study protocol has been registered with PROSPERO (ID: CRD42024518930). Moving forward, the model's generalizability should be confirmed through multicenter prospective studies.
Gong et al. (Sat,) studied this question.