A total of 142 records were identified through electronic database searches (52 records from PubMed and 90 from Embase), and 45 duplicate articles were removed. The titles and abstracts of the remaining 97 articles were assessed, of which 25 were deemed relevant. Among the relevant articles, 5 were excluded due to full-text unavailability, and 7 were considered ineligible for inclusion. The remaining 13 articles were included for the final analysis (Fig. 1).

Flow diagram of the selection process for the included studies

Figure 2 displays the QUADAS-2 evaluation of the risk of bias and applicability concerns in the 13 studies included in the systematic review.

The risk of bias and applicability concerns graph (a) and summary (b) assessed for each included article using the QUADAS-2 tool

In the patient selection domain, the risk of bias was unclear for 8 studies [14, 15, 17, 19,20,21,22,23] since they did not state whether the patients were consecutive. Similarly, the risk of bias in the index test domain was unclear in 8 studies [14, 15, 17, 19, 21, 23,24,25] as they did not indicate whether the BPE assessment was performed without knowing the pCR results. In the reference standard domain, the risk of bias was unclear in 1 study [18] owing to undefined pCR. Finally, in the flow and timing domain, it was unclear for nearly all studies since no specific period was defined between the MRI and the surgery or biopsy. In addition, 3 studies [20, 21, 23] had high applicability concerns due to the primary inclusion of patients with epidermal growth factor receptor 2 (HER2)-positive or -negative breast cancer.

Among the 13 included studies, 2 [19, 23] were multicenter studies (with 1 prospective), 1 [18] was a retrospective dual-center study, 1 [22] did not specify the study type, and the rest were retrospective single-center studies (Table 1). The sample size of the 13 studies ranged from 46 [17] to 882 [19] patients, and their mean age was between 45 and 50 years.

Significant heterogeneity was found in tumor histological type reported in 7 studies. Choi et al. [22] reported 7 histological types of breast cancer, whereas Dong et al. [20] included only HER2-positive invasive ductal carcinoma. The NACT protocol, including chemotherapeutic agents and treatment cycles, also varied considerably among the studies. In addition, tumor response assessment following chemotherapy was not uniform across the studies, and most breast cancer outcomes were categorized as pCR or non-pCR. Nonetheless, La Forgia et al. [25] and Preibsch et al. [16] assessed 4 possible outcomes according to the RECIST 1.1 criteria [28]: complete remission (rates 19% or 15%), partial response, stable disease, and progressive disease. The pCR rates fluctuated between 13.2% and 52.0% in 10 studies [14, 15, 17, 19,20,21,22,23,24, 26], and the pCR definition in these studies was not uniform.

High heterogeneity was observed across 5 aspects: field strength, manufacturer, dosage, injection rate, and DCE acquisition and sequence parameters (Table 2). Among the 13 included studies, 9 used 1.5-T MRI [14,15,16,17,18, 21, 24,25,26], 3 used 1.5-T and 3.0-T [19, 22, 23], and only 1 used 3.0-T [20]. Among the 10 articles that disclosed dosage information, the majority administered 0.2 mL/kg or 0.1 mmol/kg [15, 17, 18, 20, 23,24,25,26], while only 1 gave 0.16 mmol/kg [16] and 0.2 mmol/kg [22] each. In addition, these studies maintained the injection rate between 2 mL/s and 2.5 mL/s. Furthermore, DCE-MRI mostly contained 1 pre-contrast and 3 to 12 post-contrast acquisitions, except for 2 studies that did not report specific information [19, 23], and 1 study [17] acquired 4 pre-contrast and 7 post-contrast images. Additionally, the DCE acquisition mentioned above used different acquisition time points or temporal resolution, as observed in Table 2.

The reviewed pieces of literature were divided into 2 groups to assess the methods for BPE evaluation: studies with quantitative (n = 6; Table 3) [14, 15, 17, 19, 23, 26] or qualitative assessments (n = 8; Table 4) [16, 18, 20,21,22, 24,25,26]. For 1 article that used both methods to evaluate BPE, each was analyzed separately.

The quantitative assessment mainly involved fully automatic using fuzzy c-means clustering (n = 3) [14, 19, 26] or semi-automatic methods (n = 2) [15, 23] to segment the contralateral breast fibroglandular tissue from the region of interest (ROI) in the entire breast region. An exception was the study by Onishi et al. [19] that used the central 50% of axial images of the whole breast. The formula for the quantitative assessment varied across the studies, and the DCE phases used to assess BPE and MRI follow-up time points were also heterogeneous. The pre-contrast and first (or early) post-contrast acquisitions were the most commonly used to calculate BPE (n = 3) [15, 19, 23], with 5 out of 6 articles [14, 15, 17, 19,