Multiregional MRI-based deep learning radiomics to predict axillary response after neoadjuvant chemotherapy in breast cancer patients

TL;DR

This study developed a deep learning model using MRI data to predict if breast cancer patients will have a complete response in their axillary lymph nodes after chemotherapy, potentially avoiding unnecessary surgery.

Contribution

A novel deep learning radiomics nomogram combining intratumoral and peritumoral MRI features to predict axillary response after chemotherapy in breast cancer.

Findings

The GPTV5_DLR model achieved an average AUC of 0.876 across validation cohorts.

The DLRN model outperformed the clinical model with AUCs up to 0.958 in the training cohort.

The model showed robust performance across different patient subgroups like age and clinical stage.

Abstract

This study was designed to develop a multiregional MRI-based deep learning radiomics nomogram (DLRN) for predicting axillary pathological complete response (apCR) after neoadjuvant chemotherapy (NAC) in breast cancer.

In total, 539 patients in our hospital were randomly split into a training cohort (TC; n = 431) and an internal validation cohort (IVC; n = 108), and 703 patients were recruited from three external centers as external validation cohorts (EVC1–3). Uni- and multivariate analyses were performed to select clinicopathological characteristics and establish a clinical model. DLR models were constructed based on DL and handcrafted radiomics features extracted from gross tumor volume (GTV) and GTV incorporating 3-, 5-, 7-, and 9-mm peritumoral regions (GPTV3, GPTV5, GPTV7, and GPTV9, respectively). A DLRN model incorporating the optimal DLR model and clinicopathological predictors was developed. Model performance was assessed employing the area under the receiver operating characteristic curve (AUC), calibration curve, and decision curve analysis.

The GPTV5_DLR model surpassed the other DLR models, with an average AUC of 0.876 in the validation cohorts. The DLRN model better predicted apCR after NAC than the clinical model, demonstrating superior AUCs of 0.958 in the TC, 0.906 in the IVC, and 0.876–0.911 in EVC1–3. It also showed improved accuracy and clinical benefits for apCR prediction. Furthermore, the DLRN model achieved robust performance across different age, menstrual status, and clinical stage subgroups.

The DLRN model, based on the GPTV5_DLR model and clinicopathological features, exhibited high predictive efficiency for apCR after NAC.

The deep learning radiomics nomogram based on intra- and peritumoral regions could noninvasively predict axillary pCR in breast cancer patients receiving NAC, which might prevent patients from undergoing unnecessary axillary lymph node dissection.

Combining intratumoral and 5-mm peritumoral region radiomics had the highest predictive efficiency for axillary pCR after NAC in breast cancer.The deep learning radiomics nomogram based on intra- and peritumoral regions outperformed the clinical model.The proposed model could provide a noninvasive and easy-to-use tool to offer decision support for optimizing treatments.

Combining intratumoral and 5-mm peritumoral region radiomics had the highest predictive efficiency for axillary pCR after NAC in breast cancer.

The deep learning radiomics nomogram based on intra- and peritumoral regions outperformed the clinical model.

The proposed model could provide a noninvasive and easy-to-use tool to offer decision support for optimizing treatments.

Figures