Breast cancer (BC) is the most commonly diagnosed cancer among women, accounting for 23% of all female cancers worldwide, and its related mortality is increasing by 4% each year (1). Traditional screening methods for BC, including X-ray mammography (MMG), breast ultrasound (US), and breast magnetic resonance imaging (MRI), rely mainly on qualitative characteristics, such as the lesion density, shape of lesion margins, and enhancement pattern. These imaging methods for BC screening have limitations in the sensitivity and specificity of diagnosis. As a result, biopsies are often performed to provide a definitive diagnosis for the patient. In the era of precision medicine, improvements in the performance of BC detection are urgently needed to reduce unnecessary biopsies, which are invasive and painful. Radiomics is an emerging application that can extract innumerable quantitative image features (including descriptors of tumor shape, size, intensity, and texture) that are difficult to recognize with the naked eye from almost any medical image (2). The traditional imaging diagnostic mode is more dependent on the experience of radiologists and has strong subjectivity. Compared with traditional imaging diagnosis mode, radiomics is an emerging application that can extract innumerable quantitative image features (including descriptors of tumor shape, size, intensity, and texture) that are difficult to recognize with the naked eye from almost any medical image (2). These image features may be related to the microscopic structure and tissue biological information of tumors. Based on this, combined with clinical, pathological and genetic information, the imaging support system for clinical decision making can be constructed. A number of studies have shown that the radiomics model can improve the accuracy of breast cancer diagnosis by extracting texture features of lesion and contralateral normal breast respectively and constructing benign/malignant classifiers (3, 4). Radiomics features have proven to be of significant value in differentiating between benign and malignant breast tumors (5, 6).Therefore, radiomics provides a promising method for improving the sensitivity and specificity of the diagnosis of BCs. By refining BC detection, radiomics has the potential to reduce unnecessary invasive biopsies. In addition, Shimauchi et al. (7) found that the performance of radiologists during diagnostic tasks improved when a computer-aided diagnosis system was used. Hence, the purpose of this study was to evaluate the diagnostic efficacy of radiomics in predicting BC.

We compared the preoperative predictive value of radiomics in the diagnostic performance of BC in different studies. The results showed that the diagnostic value of radiomics was high in predicting BC with an aggregated sensitivity, specificity, and AUC of 0.84, 0.83, and 0.91, respectively. Although the number and types of features varied among the 19 included studies, which may influence the aggregated sensitivity and specificity, radiomics was shown to have good predictive ability of BC in each study. Significant heterogeneity was identified in our study. Specifically, the screening methods, selection of the scanner manufacturer and model, acquisition methods, and reconstruction parameters were shown to contribute to the heterogeneity in imaging data.

Sensitivity analyses showed that our results were reliable and stable after each study was sequentially removed, and the unreported adjusted variables were omitted.

The specificity and AUC of conventional and functional imaging analyses were similar, but conventional imaging had a slightly higher sensitivity. A possible explanation is that conventional imaging analyses included in this meta-analysis were combined with special examination methods. Li et al. combined the radiomic analysis of breast tumors and the parenchyma to improve the diagnostic accuracy of BC (19). Luo et al. used a nomogram combined with radiomics and the Breast Imaging Reporting and Data System (BI-RADS) score to predict BC (21). As for imaging modalities, the sensitivity and AUC of predicting BC by MMG were slightly higher than those of MRI and US, but the specificity was slightly lower. The use of new ultrasound imaging techniques, such as ultrasound elastography and contrast-enhanced ultrasound, may improve the detection of BC. Multi-modal MRI imaging techniques can detect most early-stage BCs, and the specificity of MRI is usually higher than that of MMG and US (31). In terms of modeling methods, ML and DL were widely studied, among which support vector machines and convolutional neural networks were the most commonly used. Logistic regression was also applied because the status of the breast mass (BC or non-BC) is a dichotomous variable. The results showed that radiomics based on either modeling method could achieve high diagnostic efficiency in predicting BC. Lastly, different data sources and study designs influenced the aggregated sensitivity, specificity, and AUC. Thus, more studies focusing on these subgroups are needed.

The preoperative diagnosis and clinical staging of BC are implemented mainly through the visual observation and analysis of medical images. The BI-RADS (32) score is a standardized description of imaging features of breast tumors, and it provides an approximate risk of malignancy to a lesion but lacks a characteristic evaluation of the intrinsic heterogeneity in tumors reflecting different biological behaviors of BC. To overcome limitations in the observation of tumor images by the naked eye, artificial intelligence has been increasingly applied to the mining and use of medical image data to meet the growing need for individualized evaluation (33). With the in-depth study of radiomics, models based on radiomics features have been shown to be a promising non-invasive method for BC classification and prediction (34). Reportedly, radiomics models based on features extracted from preoperative MMG, US, or MRI images had a relatively high predictive performance (12–30). Texture feature analysis based on US sonoelastography was first used to propose a quantitative radiomics approach for the feature selection and classification of breast tumors (28). Subsequently, many studies performed feature extraction from multi-parameter MRI images, including T2-weighted (T2w) MRI sequences, diffusion-weighted imaging (DWI) sequences, and dynamic contrast-enhanced (DCE)-MRI sequences, and constructed well-performed radiomics predictive models using ML or DL methods (12, 15, 16, 20, 23, 24, 26, 29, 30). A recent study suggested that mammography radiomics combined with quantitative three-compartment breast image analysis could reduce unnecessary breast biopsies (13).

Our meta-analysis of preoperative BC prediction using radiomics methods has two advantages. First, to the best of our knowledge, this study involving 19 articles and 5865 breast masses is the first meta-analysis to assess the diagnostic efficacy of radiomics models in predicting BC before surgery. Secondly, this study evaluated the diagnostic efficacy of radiomics models in predicting BC by comparing imaging modalities, modeling methods, and other subgroups, thereby providing ideas for subsequent radiomics research.

There are several inherent limitations to this study that need to be discussed. First, the methodology of radiomics studies included in this analysis was different as different medical centers use various examination equipment, and the selection of imaging modality, feature extraction, and modeling methods provides an infinite number of combinations. Second, the code used for feature extraction and model building was not publicly available for any of the 19 studies included in this analysis, preventing replication and independent validation of the research results. Third, because our study used summary statistics rather than individual raw data, it was not possible to achieve more reliable results. However, it was possible to achieve more precise delineation and control potential residual confounding, a common limitation of meta-analyses.

Our study shows that radiomics models based on preoperative imaging features are useful for the prediction of BC and have high diagnostic efficacy and consistency among studies. Radiomics is expected to provide a new quantitative diagnostic method for clinical work, but more well-designed prospective radiomics trials are needed to demonstrate its effectiveness and ability to translate into clinical practice.

The original contributions presented in the study are included in the article/supplementary material. Further inquiries can be directed to the corresponding author.

Guarantor of the integrity of the study: HS. Study concepts: JY. Literature research: ZL. Data acquisition: HD and YW. Data analysis: DL and ZL. Statistical analysis: HS and FZ. Manuscript preparation: ZL and YC. Manuscript editing: ZL and YC. Manuscript review: DL, HS, and JY.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.