Predicting pathological complete response to neoadjuvant chemotherapy in breast cancer patients: use of MRI radiomics data from three regions with multiple machine learning algorithms

Breast cancer (BCa) is one of the most common malignant tumors in women (Houghton and Hankinson 2021; Giaquinto et al. 2022), and neoadjuvant chemotherapy (NACT) is a common treatment for patients with locally advanced BCa. The application of chemotherapy drugs before surgery reduces the size of the tumor, and this improves the feasibility of surgical resection and patient prognosis (Early Breast Cancer Trialists' Collaborative Group (EBCTCG) 2018). However, due to the heterogeneity and complexity of BCa, some patients do not respond well to NACT. Therefore, it is crucial to predict the effect of NACT in BCa patients so that the most appropriate treatment plan can be administered.

Patients with locally advanced BCa who achieve a pathological complete response (pCR) after NACT have a good prognosis, and research has therefore focused on predicting pCR after NACT in these patients (Montemurro et al. 2020). Radiomics, a method of extracting high-dimensional data from radiological medical images using different data representation algorithms (Mayerhoefer et al. 2020), is widely used in tumor diagnosis, treatment planning, and prognostic assessment. Thus, researchers have examined the use of radiomics from magnetic resonance imaging (MRI) data as a noninvasive method for predicting pCR after NACT in BCa patients. Early radiomics studies analyzed different types of primary tumors to predict pCR (Chen et al. 2020). The use of more sophisticated radiomics methods has shown that the radiomics features extracted from peritumoral tissue can also predict which patients will achieve a pCR before treatment (Braman et al. 2017). These results characterize the environment in which the tumor grows and can also help quantify disease evolution over the next few years and possibly predict tumor recurrence or progression (Gillies et al. 2002). However, it is difficult to identify and study the tissues in which tumors may grow in future.

Therefore, Gu et al. developed the concept of the peritumoral microenvironment (Gu et al. 2022) to characterize the mechanisms of tumorigenesis and progression. The peritumor microenvironment includes tissues that surround the tumor and are in proximity to the environment where the tumor will grow. Thus, tumor occurrence, recurrence, and metastasis all depend on the peritumoral microenvironment. Although peritumoral tissue includes regions with tumor cell infiltration and non-infiltrated normal tissue, the peritumoral microenvironment refers to non-tumor tissue that provides a suitable environment for tumor cell growth. Therefore, some studies of BCa have focused on background parenchymal enhancement (BPE) in breast magnetic resonance imaging (MRI) because it is more representative of the peritumor microenvironment (Wang et al. 2015). BPE is defined as a normal background enhancement of fibroglandular tissue after gadolinium injection (Rella et al. 2018; Telegrafo et al. 2016a) and is associated with physiological vascularization and perfusion of breast tissue (Wu et al. 2017; Brooks et al. 2018). There is evidence that a high BPE is associated with tumor malignancy, lymph node metastasis, and increased risk of recurrence (Telegrafo et al. 2016b). There is also evidence that a decrease in the BPE after chemotherapy is associated with a favorable response to NACT (Rella et al. 2020; Preibsch et al. 2016; La Forgia et al. 2021). Therefore, we hypothesized that the use of radiomic analysis of BPE tissues for characterization of the tumor and peritumoral microenvironment would provide improved predictions of pCR in women who receive NACT for locally advanced BCa.

This study examined the value of MRI radiomics features extracted from intratumoral, peritumoral, and BPE for prediction of pCR after NACT in patients with locally advanced BCa and also determined whether the combination of different tissue regions, especially the peritumoral microenvironment, can improve these predictions.

Our major results are that use of a radiomics signature based on the BPE region provided better predictions of pCR than signatures from the intratumoral region and the peritumoral region, and that addition of the BPE region to the intratumoral and the peritumoral regions improved predictions of the XGBoost model. These results demonstrate that the peritumoral microenvironment, as indicated by BPE, has greater prognostic value than that of the tumor parenchymal area. In addition, a radiomics model based on the combined use of the intratumoral, peritumoral, and BPE regions provided reliable predictions of pCR when using the XGBoost machine learning model. These results confirm the importance of machine learning for prediction of pCR for women with invasive BCa without distant metastases who completed a standard NACT regimen (Radak et al. 2023).

MRI, especially functional imaging such as dynamic contrast-enhanced MRI (DCE-MRI) and diffusion weighted imaging (Curigliano et al. 2017), is playing an increasingly important role in the diagnosis and treatment of BCa. In particular, clinicians use these imaging results to evaluate the morphology, signal strength, and enhancement of the tumor area (Negrão et al. 2019; Lother et al. 2023; Calabrese et al. 2021). However, these classical indicators do not fully reflect the biological behavior of lesions, and they only have limited use for evaluations of treatment response. Therefore, a key direction of recent research has been to obtain more biological information from multi-level and multi-angle omics data to achieve more accurate predictions of tumor status and response to treatment. Radiomics is a new omics method that uses medical imaging results with different algorithms as a non-invasive method to evaluate response to treatment and predict treatment outcome in patients with BCa (Conti et al. 2021). The results of the present study confirmed the usefulness of a radiomics model that combines data from the intratumoral, peritumoral, and BPE regions for prediction of pCR in BCa patients. We also found that the use of all three of these regions provided better predictions than any single region alone. This suggests that the jointly constructed signature provides more comprehensive information and more accurately reflects the biological characteristics and microenvironment of these tumors.

Compared with traditional imaging studies, our radiomics model provided more objective, accurate, and comprehensive information for predicting pCR in BCa patients who received NACT (Liu et al. 2019), and this improvement was due to our use of a combination of radiomics signatures of the peritumoral and BPE regions. Previous studies showed that the radiomics features of the peritumoral region reflected the response of surrounding tissues, such as inflammation and fibrosis. Therefore, the features of the peritumoral region have high specificity in predicting pCR in patients with BCa (Hussain et al. 2021). Our results are therefore consistent, in that the predictive performance of the peritumoral region was greater than that of the intratumoral region. Our novel finding is that the predictive value of the BPE was higher than that of the peritumoral region, presumably due to the unique characteristics of the BPE. The BPE represents the peritumoral microenvironment, and this signal is related to microvessel density and local blood flow in the breast tissue. Thus, BPE may have a high predictive value because a greater enhancement represents an environment that is more favorable for tumor development, growth, and metastasis (Liao et al. 2020). Similar studies have confirmed that BPE is an important predictor of neoadjuvant chemotherapy response. For example, Vignesh A Arasu et al. combined FVT and BPE to predict pCR. The associated BPE area under the curve (Griethuysen et al. 2017) was 0.77 (95% CI 0.56–0.98) (Arasu et al. 2020). This is similar to our result of 0.777 (95% CI 0.672, 0.866). However, predictive performance has no substantial additive improvement when adding BPE to an FTV model. On the contrary, in our study, the diagnostic performance of the intratumoral and peritumoral regions improved after adding the BPE signature. This may be because radiomics features better reflect the heterogeneity of the BPE region. In addition, we also found that radiomics signatures in the tumor region had the lowest predictive value. This may be because the complexity and heterogeneity of tumor tissues lead to unstable predictions. However, we also found that adding the BPE region and the peritumoral region in the joint model significantly improved predictions, and that the BPE region had a greater impact than the peritumoral region. This demonstrates that the peritumoral microenvironment contains more information related to tumor occurrence, development, and prognosis after treatment. Our SHAP analysis also showed that the most important features of the prediction model were from the peritumoral and BPE regions, consistent with our finding that the individual diagnostic value of the peritumoral region and the BPE region was higher than that of the intratumoral region. Our finding that pairwise combinations of radiomics signatures had greater predictive values than those of individual signatures confirmed the importance of using all the three regions for making the best predictions of pCR.

MRI-based radiomics has been widely used for prediction of pCR in patients with BCa (O'Donnell et al. 2022). Compared with previous studies, our approach has certain advantages. Guo et al. used the radiomics features of DCE-MRI combined with clinical features to predict pCR and reported an AUC value of 0.864 in the test set (Guo et al. 2022), similar to our AUC values (training set: 0.891, validation set: 0.861). However, the performance of this previous prediction model from the sole use of imaging omics features was only 0.842. Our results were better, possibly because we tested a variety of machine learning algorithms, and used the optimal algorithm—XGBoost—to construct a prediction model. Similarly, Tahmassasebi et al. (2019) used a variety of machine learning models to predict pCR in patients with BCa and showed that the XGBoost classifier provided the most stable performance and highest accuracy, similar to our results. Umutlu et al. developed a prediction model for pCR in patients with BCa and reported an AUC of 0.92 (Umutlu et al. 2022), higher than that in our study. However, this previous study used multimodal radiomics features from the fusion of images derived from PET and MRI, whereas we only used MRI results. Thus, when considering costs and simplicity, our approach may be more suitable for clinical practice.

In conclusion, this study used radiomics data from three different regions of breast tissue and six machine learning algorithms to predict pCR in patients with BCa. Our results provide a theoretical basis for the peritumoral microenvironment affecting the efficacy of NACT for these patients. We suggest that future research should explore the biological mechanism by which the peritumoral microenvironment affects the response to NACT, because this may facilitate the identification of new therapeutic targets for BCa and the development of more individualized treatments.