Triple-negative breast cancer (TNBC) accounts for approximately 10% to 20% of breast cancers but almost 30% to 40% of breast cancer–related deaths (1). TNBC is defined by a profile of negative immunohistochemical staining of the receptors for progesterone, estrogen, and human epidermal growth factor (HER2) and therefore not responsive to endocrine or HER2-targeted therapies (2). In addition, TNBC, when not responsive to chemotherapy, is generally associated with a poor prognosis with a high recurrence rate and a low long-term survival rate (3).

In patients with TNBC, neoadjuvant systemic therapy (NAST) with chemotherapy agents such as anthracyclines, taxanes, and cyclophosphamide, carboplatin plus Food and Drug Administration–approved immunotherapy agents such as pembrolizumab is usually administered before surgery to downstage the tumor (2, 4–8). Patients with a pathologic complete response (pCR) to NAST have favorable long-term overall survival and event-free survival, whereas patients without a pCR to NAST have higher recurrence and mortality rates (9, 10). However, only 50% to 60% of TNBC patients achieve a pCR to NAST. Thus, early prediction of NAST response is crucial to avoid exposure of predicted non-responders to ineffective NAST and unnecessary toxicity of neoadjuvant immunotherapy. Early prediction of NAST response can also help oncologists triage patients to a clinical trial and has the potential to better personalize therapy.

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) and the associated temporal enhancement curves can be used in pharmacokinetic modeling to determine the vascular properties of a tumor, such as the integrity and density of the tumor micro-vessels (11). Volumetric changes along with quantitative and semiquantitative kinetic parameters derived from DCE-MRI have been found to be applicable for breast cancer diagnosis, measurement of breast tumor size, classification of breast tumors, and detection of residual breast cancer after NAST (12–18).

In recent years, several studies have been conducted to assess the role of quantitative radiomic imaging features extracted from MRI images in predicting prognosis and treatment response in patients with breast cancer (19, 20). Wu et al. (21) studied quantitative radiomic features from DCE-MRI extracted before and after one cycle of NAST and demonstrated that these features could be used to develop a clinical biomarker for the prediction of the tumor response to NAST. Other investigators also studied radiomic features from DCE-MRI images for the early prediction of tumor response to NAST (12, 19, 21–24). In addition, radiomic features of breast MRI images have been widely investigated for the noninvasive characterization of tumors (25–28).

Studies suggest that peritumoral features assessed by MRI could provide vital information on the tumor microenvironment, cancer development, chemoresistance, and treatment response (29, 30). While previous studies have shown the potential of radiomic analysis in predicting response to NAST, most of these studies have focused on the tumoral region alone. Furthermore, few studies have explored the performance of radiomic models, specifically in patients with TNBC, and even fewer have evaluated the performance of models based on features extracted from DCE-MRI images obtained at multiple time points during NAST.

In this study, we systematically evaluated the performance of models based on radiomic features of both the peritumoral and tumoral regions from DCE-MRI images at baseline (BL), after two cycles (C2), and after four cycles (C4) of NAST for early prediction of response to NAST in TNBC.

In this study, we used longitudinal DCE-MRI images obtained before the start of doxorubicin and cyclophosphamide-based NAST and after two and four cycles of NAST to assess radiomic features from the peritumoral and tumoral regions for early prediction of NAST response in TNBC patients. In the univariate analysis, we identified 46 radiomic features able to predict pCR with an AUC at least 0.70 for both the training and testing cohorts. Furthermore, in the multivariate analysis, we found that 13 multivariate radiomic models had AUC at least 0.75 for both the training and testing cohorts for early prediction of NAST response in TNBC.

Our results revealed that FO radiomic features from DCE-MRI were better predictors of treatment response than GLCM features. None of the 300 GLCM features had AUC at least 0.70 for both the training and testing cohorts; thus, their usefulness in predicting NAST response could not be established. The interreader variability analysis showed high reliability and reproducibility with a slightly better interreader agreement in GLCM features than in FO features. GLCM features capture texture information of the image, which is less subjective and more reproducible across readers than the FO features, which are mostly visual features. Furthermore, the features extracted from baseline DCE-MRI images and the models using features only from baseline DCE-MRI images showed poor performance in terms of AUC. Similar to this finding, Panthi et al. have previously reported that the tumor size measurements (longest dimension, tumor volume, enhanced tumor volume and functional tumor volume) extracted from DCE-MRI images at baseline showed poor performance. The tumor measurements extracted at C2, C4, and their relative differences (C2 vs BL and C4 vs BL) showed good correlation with the treatment response with a maximum AUC of 0.84 [95% CI: 0.76-0.92] for functional tumor volume at C4 (31). The possible explanation of these findings is that tumor biology can evolve over the course of treatment. The radiomic features measured after the initiation of NAST may capture these changes in tumor biology better than features measured at baseline, contributing to the better predictive performance of the features measured after NAST initiation.

In our study, almost 50% (6/13) of the radiomic models with AUC at least 0.75 for both the training and testing cohorts were based on a combination of peritumoral and tumoral features. The inclusion of the peritumoral features along with the tumoral features may have improved the performance of radiomic models by capturing the infiltrative tumor margins and information about the tumor microenvironment. Additionally, peritumoral features have the potential to aid in predicting treatment response, as tumor behavior heterogeneity likely reflects variation both of the tumor and surrounding environment.

MRI-based radiomic analysis has shown promise in predicting treatment response in breast cancer (12, 19, 22, 23, 34). Cain et al. used pretreatment tumor features from 151 patients with TNBC and HER2-enriched breast cancer in a multivariate machine learning model and showed an AUC of 0.707 for pCR prediction (35). In a study including 83 breast cancer patients, Pesapane et al. extracted 136 representative radiomic features of the tumoral region from pretreatment T1-weighted contrast-enhanced MRI images to predict the response to NAST. Their radiomic model had an AUC of 0.64 (95% CI, 0.51-0.75), which increased to 0.83 (95% CI, 0.73-0.92) after the investigators combined the radiomic model with clinical and biological features (36). Similarly, Fan et al. performed a radiomic analysis of pretreatment DCE-MRI images in 57 patients and demonstrated that combining features from the tumoral region and background parenchyma significantly improved the performance of their radiomic model (19). Li et al, in a study of 33 patients, compared AUCs for pretreatment images and images obtained after one cycle of NAST and observed better predictive AUCs after one cycle of treatment (37). In contrast to the aforementioned studies, we created multivariate radiomic models based on DCE-MRI images at three imaging time points to extract features from peritumoral as well as tumoral regions in a much larger cohort limited to patients with TNBC.

Braman et al. conducted a study including 117 breast cancer patients (with hormone-receptor-positive, HER2-negative, triple-negative, and HER2-positive disease) and showed that combined peritumoral and tumoral radiomic features obtained at baseline could be utilized to predict pCR to neoadjuvant chemotherapy (22). These authors reported an AUC of 0.74 for the testing cohort, which was slightly lower than the AUC of 0.79 for the testing cohort in our study, which had a larger patient pool, was limited to patients with TNBC, and was based on radiomic features obtained at multiple time points. Caballo et al. (38) used a four-dimensional machine learning radiomics approach to assess breast cancer response to NAST. They extracted 348 features from peritumoral and tumoral regions from DCE-MRI images to develop predictive radiomic models. These authors reported a multivariate model with an AUC of 0.71 for the data set of 251 patients (107 with luminal A subtype, 47 with luminal B subtype, 25 with HER2-enriched subtype, and 72 with TNBC). Furthermore, their study showed that the predictive performance improved when the radiomic models were tailored to specific subtypes of breast cancer. Caballo et al. found an AUC of 0.80 in 72 TNBC patients (38), which is comparable to the testing-group AUC of 0.79 in our study of 163 TNBC patients.

Fan et al. reported a study in which clinical features and radiomic features from the tumoral region before and after two cycles of treatment were acquired for 114 patients with primary breast cancer. These authors observed an AUC of 0.57 for the models based on pretreatment features and an AUC of 0.77 for the model based on posttreatment features. The combined model based on pretreatment features, posttreatment features, relative net feature change, Jacobian maps, and clinical features had an AUC of 0.81 (39). Our findings are consistent with their findings in that our findings showed improvement in the predictive performance with a combined model with features from BL, C2, C4, and their net feature changes.

Our study has several limitations. First, our study was conducted on a single scanner platform at a single institution, and therefore our findings will need to be validated on other scanner platforms and at different institutions. Second, although our study is the largest radiomic analysis reported to date for treatment response prediction in TNBC patients, additional validation may be necessary in a larger study with better statistical power and was only tested on patients with preoperative AC chemotherapy. Finally, the peritumoral regions were uniformly defined with a fixed dilation of 10 mm around the tumor ROIs for all patients, regardless of tumor size. Future studies with different dilation radii based on individual tumor properties can offer a more comprehensive understanding of the role of the peritumoral region in response prediction.

In summary, our study shows that radiomic models based on peritumoral and tumoral features from longitudinal DCE-MRI images were able to serve as noninvasive biomarkers for early prediction of NAST response in patients with early-stage TNBC. Radiomic features extracted after four cycles of treatment and their change relative to baseline were better predictors of response than those from baseline only. Further, FO radiomic features showed better predictive performance than GLCM features.

The datasets presented in this article are not readily available because we are not allowed to share patient data. Requests to access the datasets should be directed to jma@mdanderson.org.

The studies involving humans were approved by The UT MD Anderson Cancer Center IRB. The studies were conducted in accordance with the local legislation and institutional requirements. Written informed consent for participation in this study was provided by the participants’ legal guardians/next of kin.

BP: Conceptualization, Formal Analysis, Methodology, Writing – original draft. RMM: Methodology, Writing – review & editing. BA: Data curation, Methodology, Writing – review & editing. MB: Formal Analysis, Methodology, Writing – review & editing. RC: Data curation, Methodology, Writing – review & editing. HC: Formal Analysis, Writing – review & editing. KKH: Writing – review & editing. LH: Writing – review & editing. K-PH: Writing – review & editing. AK: Writing – review & editing. DL: Writing – review & editing. HL-P: Writing – review & editing. JWTL: Writing – review & editing. JKL: Writing – review & editing. SP: Formal Analysis, Methodology, Writing – review & editing. FP: Writing – review & editing. JBS: Formal Analysis, Methodology, Writing – review & editing. JS: Formal Analysis, Writing – review & editing. AT: Writing – review & editing. DT: Writing – review & editing. VV: Writing – review & editing. PW: Writing – review & editing. JW: Writing – review & editing. ZX: Formal Analysis, Methodology, Writing – review & editing. WY: Writing – review & editing. ZZ: Formal Analysis, Methodology, Writing – review & editing. CY: Writing – review & editing. GR: Conceptualization, Formal Analysis, Methodology, Supervision, Writing – review & editing. JM: Formal Analysis, Methodology, Supervision, Writing – review & editing.