Deep learning radiomics of ultrasonography for comprehensively predicting tumor and axillary lymph node status after neoadjuvant chemotherapy in breast cancer patients: A multicenter study

Abstract

Background: Neoadjuvant chemotherapy (NAC) can downstage tumors and axillary lymph nodes in breast cancer (BC) patients. However, tumors and axillary response to NAC are not parallel and vary among patients. This study aims to explore the feasibility of deep learning radiomics nomogram (DLRN) for independently predicting the status of tumors and lymph node metastasis (LNM) after NAC.

Methods: In total, 484 BC patients who completed NAC from two hospitals (H1: 297 patients in the training cohort and 99 patients in the validation cohort; H2: 88 patients in the test cohort) were retrospectively enrolled. The authors developed two deep learning radiomics (DLR) models for personalized prediction of the tumor pathologic complete response (PCR) to NAC (DLR-PCR) and the LNM status (DLR-LNM) after NAC based on pre-NAC and after-NAC ultrasonography images. Furthermore, they proposed two DLRNs (DLRN-PCR and DLRN-LNM) for two different tasks based on the clinical characteristics and DLR scores, which were generated from both DLR-PCR and DLR-LNM.

Results: In the validation and test cohorts, DLRN-PCR exhibited areas under the receiver operating characteristic curves (AUCs) of 0.903 and 0.896 with sensitivities of 91.2% and 75.0%, respectively. DLRN-LNM achieved AUCs of 0.853 and 0.863, specificities of 82.0% and 81.8%, and negative predictive values of 81.3% and 87.2% in the validation and test cohorts, respectively. The two DLRN models achieved satisfactory predictive performance based on different BC subtypes.

Conclusions: The proposed DLRN models have the potential to accurately predict the tumor PCR and LNM status after NAC.

Plain language summary: In this study, we proposed two deep learning radiomics nomogram models based on pre-neoadjuvant chemotherapy (NAC) and preoperative ultrasonography images for independently predicting the status of tumor and axillary lymph node (ALN) after NAC. A more comprehensive assessment of the patient's condition after NAC can be achieved by predicting the status of the tumor and ALN separately. Our model can potentially provide a noninvasive and personalized method to offer decision support for organ preservation and avoidance of excessive surgery.

Keywords: breast cancer; deep learning; lymph node metastasis; neoadjuvant chemotherapy; pathologic complete response; treatment decision; ultrasonography.