Multi-modal radiomics model based on four imaging modalities for predicting pathological complete response to neoadjuvant treatment in breast cancer.

OBJECTIVE

The radiomics model based on single imaging modality has been demonstrated as a promising approach for predicting the response to neoadjuvant treatment (NAT) in breast cancer. However, whether integrating multiple imaging modalities improve the performance of the radiomics model is undetermined. This study aims to develop a multi-modal radiomics model based on four imaging modalities, including ultrasound (US), mammography (MM), computed tomography (CT), and magnetic resonance imaging (MRI), for predicting pathological complete response (pCR) in breast cancer after NAT.

METHODS

Patients who underwent surgery after NAT from January 2019 to July 2023 were retrospectively studied. Univariate and multivariate analyses were performed to identify independent clinical risk factors for pCR. The radiomic features were extracted from the volume of interest on the four imaging modalities. The least absolute shrinkage and selection operator was used for developing radiomic signatures. The multi-modal radiomics model was developed by combining four radiomic signatures. The combined model was developed by combining clinical risk factors and four radiomic signatures. A nomogram was developed to visualize the combined model. Model performance was internally validated by using the five-fold cross-validation.

RESULTS

In total, 89 patients were included, with the pCR rate of 31.5% (28/89). Multivariate analyses identified PR status (OR = 4.450, 95% confidence interval [CI], 1.228-18.063, P = 0.028), HER2 status (OR = 9.95, 95% CI, 1.525-201.894, P = 0.044) and clinical T stage (OR = 0.253, 95% CI, 0.076-0.753, P = 0.016) were independent clinical risk factors for pCR. The AUCs and brier scores of the radiomic signatures of US, MM, CT, and MRI were 0.702 (95% CI: 0.583-0.821), 0.762 (95% CI: 0.660-0.865), 0.814 (95% CI: 0.725-0.903), 0.787 (95% CI: 0.685-0.889) and 0.198, 0.177, 0.165, 0.170 respectively. The performance of the multi-modal radiomics model was superior to all radiomic signatures with an AUC of 0.904 (95% CI: 0.838-0.970) and with the brier score of 0.111. After adding independent clinical risk factors, the performance of the combined model further improved, achieving an AUC of 0.943 (95% CI: 0.893-0.992) and a brier score of 0.082. The nomogram showed potential clinical value.

CONCLUSION

The multi-modal radiomics model based on US, MM, CT, and MRI could accurately predict pCR in breast cancer after NAT, which was superior to all radiomic signatures. Incorporating clinical risk factors may further improve the performance of the muti-modal radiomics model, which could provide valuable information for guiding treatment decisions.