Triple-negative breast cancer (TNBC) is a clinical challenge because of its invasive nature, high risk of distant metastasis, and poor prognosis. Compared with other breast cancer patients, TNBC patients are 2–3.5 times more likely to have distant recurrence (Fatayer et al., 2016). It has been demonstrated that the probability of a pathological complete response (pCR) is higher in TNBC patients who receive neoadjuvant therapy (NAC) (close to 31% at present) than in patients with other molecular subtypes, suggesting that NAC improves DFS in this group of patients (Houssami et al., 2012). However, pCR alone is not enough to predict the long-term recurrence-free survival rate of patients with TNBC, and an efficient prognostic biomarker is urgently needed to help stratify patients and create treatment guidelines.

Recently, some studies have indicated that radiomics can be used to obtain a series of related parameters to quantify the heterogeneity of lesions and shows promise for improving tumour prognosis. In previous studies, the radiomics nomogram provided a promising prediction of neoadjuvant chemotherapy efficacy in breast cancer patients based on pretreatment MRI images (Bian et al., 2020; Chen et al., 2020). Another study reported that the radiomics signature(Rad-score) could be used for DFS prediction in HER-2-positive invasive breast cancer treated with NAC, and the radiomics-clinicoradiologic-based nomogram may potentially be useful for personalized treatment strategies (Li et al., 2020). However, there is no relevant research on TNBC.

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) has excellent sensitivity and good specificity for breast cancer diagnosis and plays an important role in characterizing the heterogeneity of tumours. Most studies involving radiomics analysis only use the initial enhancement phase of DCE-MRI, and the additional value of radiomics calculated from later enhancement images was limited. Nevertheless, the radiomics features derived from the phases of multiple DCE-MRI images cannot be ignored, which may imply more information changing over time points.

The purpose of this study was to investigate whether the radiomics derived from all DCE-MRI phases obtained in TNBC patients before NAC are associated with DFS and to compare the combined radiomics nomogram and the clinicoradiological nomogram for their abilities in predicting DFS in patients with TNBC treated with NAC.

The institutional review board approved this two-institution study and retrospective radiomics data analysis (approval No: 2004216-14), and the requirement for written informed consent was waived.

We compared patient characteristics using commercially available statistical software (IBM SPSS 24.0). When appropriate, significant differences between the training and validation groups were assessed by the Chi-square test, Fisher’s test or t-test. A two-sided p value of less than 0.05 indicates a significant difference. The Rad-scores were divided into two groups (high-risk vs low-risk) using receiver operating characteristic (ROC) curve analysis according to optimal cut-off value determined by maximizing the Youden index (sensitivity + specificity-1). Significant variables in the univariate Cox proportional hazard model (p < 0.05) were included in the multivariate analysis. The combined radiomics nomogram incorporated the radiomics signature and various independent risk factors based on multivariate analysis in the training group and was then validated in the validation group. The predictive ability and discriminatory performance of each established model were evaluated using an index of probability of concordance (C-index), and the C-index between the predicted probability and actual outcome was calculated to evaluate the predictive ability and discrimination of the model (Wolbers et al., 2009). The value of the C-index ranges from 0.5–1.0, with 0.5 indicating random chance and 1.0 indicating perfectly accurate discrimination. The nomograms were subjected to bootstrapping validation (1000 bootstrap resamples) to obtain a relatively corrected C-index.

In our study, we demonstrated the prognostic value of multiphase CE-MRI radiomics features for patients with TNBC treated with NAC. In addition, we developed a combined radiomics model that incorporates the radiomics signature and MRI and pathology findings for the individualized prediction of DFS in TNBC patients who underwent NAC. Compared with the clinicoradiological nomogram, the combined radiomics nomogram had superior prognostic performance in DFS estimation.

For feature extraction and selection, we measured ten new sequential features to characterize the textural changes observed on DCE images over time series. These features have not previously been used or described in the domain of breast radiomics except in Li et al. (2020) and Xie et al. (2019) studies, who used new features to differentiate different subtypes of breast cancer and predict DFS in patients with HER2-positive breast cancer treated with NAC. We compared two models to investigate whether the accuracy of the radiomics model was significantly improved after adding new features. Roberto et al. (Lo Gullo et al., 2020) and Gibbs et al. (2019) both demonstrated that delayed postcontrast phases did not add any significant discriminative value to the analysis, which is inconsistent with our research results. The reason may be that we added new sequential features, but they did not include them, and the subjects of their study were subcentimetre masses that were much smaller than ours lesions. Furthermore, the sequential texture features derived from dynamic phases may capture information on both spatial heterogeneity and tumour perfusion, which is more valuable in predicting DFS than differentiating benign and malignant lesions.

In our study, the final Rad-score calculation formula included six potential features all from the new sequential features. The six selected radiomics features comprised one from skewness, two from entropy, one from Kendall-tau-b and two from conservation. Among them, other studies have also emphasized the importance of skewness and entropy in reflecting the heterogeneity of tumours. Kendall-tau-b and conservation were calculated from interactive information between the current subject and the remainder of the subjects, which means that if the changes increased, the Rad-score increased, indicating a worse prognosis. One possible interpretation is that this change may be related to the high perfusion of the tumours, and tumours with abundant blood supply tend to be more heterogeneous and have a worse prognosis. Attentionally, three of the six selected features were GLCM (grey level cooccurrence matrix), and two were NGTDM (neighbourhood grey-tone difference matrix). At present, GLCM is the most widely used texture extraction method, which has also been confirmed in assessing tumour heterogeneity and plays a very important role in various fields. The basic principle of the GLCM is based on spatial correlation between neighbouring pixels. NGTDM represents contrast, which is determined by changes in intensity between a target voxel and the surrounding neighbours and then enables the calculation of the apparent difference between neighbouring regions of voxel intensities. Contrast is also related to tumour heterogeneity; tumours with poor prognosis tend to have higher contrast (Sun and Wee, 1983). Our results also showed that the Rad-score had a promising high value for predicting DFS, which was confirmed by Kaplan–Meier survival curves in the training group (p < 0.0001) and in the validation group (p < 0.0001). Interestingly, the cut-off value (Rad-score = 0.2528. for predicting DFS was similar to QL’s study (Rad-score = 0.2523), regardless of TNBC or HER2-positive breast cancer with NAC.

There were differences in clinical stage, pre-NAC N stage and pCR status between the training and validation groups, which might be associated with differences in study populations with different hospitals. In the validation group, the later the clinical stage, the more difficult it was to achieve pCR. Various previous studies have confirmed that a tumour’s response to neoadjuvant therapy provides prognostic information. The attainment of a pCR after NAC and surgical resection improved the DFS rate of patients (Houssami et al., 2012), (Cortazar et al., 2014; Chen et al., 2017; Symmans et al., 2017), consistent with our study. However, 42.2% of the patients in the training group received pCR after NAC, and this rate is higher than those reported in other studies (Houssami et al., 2012), potentially because we ruled out patients who did not undergo surgery and did not finish a complete NAC regimen. Interestingly, in the training group, multifocal/centric lesions were identified as independent predictors for the DFS of TNBC after NAC. Many studies (Duraker and Çaynak, 2014; Lang et al., 2017) have demonstrated that multifocal/centric foci exhibit more biologically aggressive behaviour than has been observed for unifocal breast cancer, and this could influence DFS and OS. Although the multifocal/centric lexicon was not included in BI-RADS, these patients should receive more attention during postoperative follow-up. While Park et al. (2018a) found that N-stage was a predictor of DFS in breast cancer, our analysis failed to support these findings, possibly due to differences in study populations. In addition, the features at MR imaging(mass vs nonmass) was not associated with DFS in multivariate analysis of variance in our study, which was consistent with the study of Tahmassebi et al. (2019).

The prognostic ability of radiomics signatures has been demonstrated in many studies. For example, Li et al. (2016) suggested that image-based radiomics features may be helpful in assessing the risk of breast cancer recurrence. Park et al. (2018b) demonstrated that Rad-scores generated from radiomics signatures based on preoperative MRI have prognostic value. In our study, we analysed preoperative MRI findings in TNBC, a special pathological type of breast cancer, and supported the notion that the Rad-score helps stratify patients, and patients from high-risk groups need more careful follow-up management.

In this study, we developed a radiomics signature-based nomogram for the individualized prediction of recurrence in patients with TNBC after NAC. The nomogram incorporates three components of a radiomics signature with six selected features, including pCR status and MR findings indicating multifocal/centric lesions, which is promising to facilitate individualized predictions and the prediction of follow-up needs in patients with poor outcomes with regard to DFS.

Our study has several limitations. First, this is a retrospective study. Second, most of the patients were examined using MR after a biopsy, which might have affected assessments. Third, we discuss only DCE images in our study, and further prospective studies should include a variety of breast MR imaging protocols, such as T2W, DWI, and DCE-MRI.

In conclusion, the results of our study show that the identified Rad-score has the potential to be used as a biomarker for risk stratification for DFS in patients with TNBC after NAC. In addition, our results show that a radiomics nomogram that incorporates a radiomics signature and MRI and clinicopathological findings can be used to facilitate the individualized prediction of recurrence in patients with TNBC after NAC and surgery. This type of quantitative radiomics prognostic model of breast cancer could be useful for precision medicine and could affect patient follow-up strategies.