Longitudinal MRI-based fusion novel model predicts pathological complete response in breast cancer treated with neoadjuvant chemotherapy: a multicenter, retrospective study

Abstract

Background: Accurate identification of pCR to neoadjuvant chemotherapy (NAC) is essential for determining appropriate surgery strategy and guiding resection extent in breast cancer. However, a non-invasive tool to predict pCR accurately is lacking. Our study aims to develop ensemble learning models using longitudinal multiparametric MRI to predict pCR in breast cancer.

Methods: From July 2015 to December 2021, we collected pre-NAC and post-NAC multiparametric MRI sequences per patient. We then extracted 14,676 radiomics and 4096 deep learning features and calculated additional delta-value features. In the primary cohort (n = 409), the inter-class correlation coefficient test, U-test, Boruta and the least absolute shrinkage and selection operator regression were used to select the most significant features for each subtype of breast cancer. Five machine learning classifiers were then developed to predict pCR accurately for each subtype. The ensemble learning strategy was used to integrate the single-modality models. The diagnostic performances of models were evaluated in the three external cohorts (n = 343, 170 and 340, respectively).

Findings: A total of 1262 patients with breast cancer from four centers were enrolled in this study, and pCR rates were 10.6% (52/491), 54.3% (323/595) and 37.5% (66/176) in HR+/HER2-, HER2+ and TNBC subtype, respectively. Finally, 20, 15 and 13 features were selected to construct the machine learning models in HR+/HER2-, HER2+ and TNBC subtypes, respectively. The multi-Layer Perception (MLP) yields the best diagnostic performances in all subtypes. For the three subtypes, the stacking model integrating pre-, post- and delta-models yielded the highest AUCs of 0.959, 0.974 and 0.958 in the primary cohort, and AUCs of 0.882-0.908, 0.896-0.929 and 0.837-0.901 in the external validation cohorts, respectively. The stacking model had accuracies of 85.0%-88.9%, sensitivities of 80.0%-86.3%, and specificities of 87.4%-91.5% in the external validation cohorts.

Interpretation: Our study established a novel tool to predict the responses of breast cancer to NAC and achieve excellent performance. The models could help to determine post-NAC surgery strategy for breast cancer.

Funding: This study is supported by grants from the National Natural Science Foundation of China (82171898, 82103093), the Deng Feng project of high-level hospital construction (DFJHBF202109), the Guangdong Basic and Applied Basic Research Foundation (grant number, 2020A1515010346, 2022A1515012277), the Science and Technology Planning Project of Guangzhou City (202002030236), the Beijing Medical Award Foundation (YXJL-2020-0941-0758), and the Beijing Science and Technology Innovation Medical Development Foundation (KC2022-ZZ-0091-5). Funding sources were not involved in the study design, data collection, analysis and interpretation, writing of the report, or decision to submit the article for publication.

Keywords: Breast cancer; Deep learning; Longitudinal radiomics; Multi-parametric MRI; Neoadjuvant chemotherapy; Pathological complete response.

Fig. 1

The study design and workflow of longitudinal MRI-based radiomics deep learning in predicting pCR to neoadjuvant chemotherapy.

Fig. 2

The Spearman correlation coefficient network diagrams showed the relations between each pair of selected features in HR+/HER2− (a), HER2+ (c) and TNBC subtype (e). the Spearman correlation coefficient heat maps showed the relations between selected features and clinical characteristics in HR+/HER2− (b), HER2+ (d) and TNBC subtype (f). Each feature was independent predictor as there was no correlation coefficient >0.8 in each subtype. And all the imaging-derived features were independent from clinical characteristics as no correlation coefficient >0.8 was observed in each subtype.

Fig. 3

The horizontal bar charts showed the feature importance of the selected radiomics and deep learning features in HR+/HER2− (a), HER2+ (b) and TNBC subtype (c). In the three random forest models, all the Shapley values of selected features were higher than the corresponding max shadow value in each subtype. It indicated that all the features contributed to develop the models.

Fig. 4

Predictive performances of the different models in the primary and external validation cohorts (a–l). Plots show the ROC curves of stacking model, pre-model, post-model and delta-model, in HR+/HER2− (a), HER2+ (b) and TNBC subtype (c), respectively, in the primary cohort. Plots show the ROC curves of stacking model, pre-model, post-model and delta-model, in HR+/HER2− (d), HER2+ (e) and TNBC subtype (f), respectively, in the validation cohort 1. Plots show the ROC curves of stacking model, pre-model, post-model and delta-model, in HR+/HER2− (g), HER2+ (h) and TNBC subtype (i), respectively, in the validation cohort 2. Plots show the ROC curves of the stacking model, pre-model, post-model and delta-model, in HR+/HER2− (j), HER2+ (k) and TNBC subtype (l), respectively, in the validation cohort 3.

Fig. 5

Predictive performances of the different models in the primary and external validation cohorts (a–f). Plots show the decision curves of stacking model, pre-model, post-model and delta-model, in HR+/HER2− (a), HER2+ (c) and TNBC subtype (e), respectively, in the primary cohort. Plots show the decision curves of the stacking model, pre-model, post-model and delta-model, in HR+/HER2− (b), HER2+ (d) and TNBC subtype (f), respectively, in the validation cohorts.