Predicting breast cancer pathological complete response with clinical and imaging data

Area of Science:

• Oncology
• Radiology
• Machine Learning

Background:

• Pathological complete response (pCR) is a critical prognostic factor for breast cancer (BC) patients treated with neoadjuvant chemotherapy (NAC).
• Existing unimodal prediction models for pCR have limitations.
• Multimodal machine learning approaches are needed to improve prediction accuracy.

Purpose of the Study:

• To develop and evaluate a combined clinical-radiomics machine learning model for predicting pCR in BC patients receiving NAC.
• To compare the performance of the combined model against unimodal clinical and radiomics models.

Main Methods:

• Retrospective analysis of 211 BC patients.
• Development of a radiomics score (Radscore) from multiparametric MRI data.
• Integration of Radscore with clinical predictors using machine learning to create clinical, radiomics, and combined models.

Main Results:

• The combined clinical-radiomics model demonstrated superior performance compared to unimodal models.
• The combined model achieved an AUC of 0.937 in the training cohort and 0.853 in the validation cohort.
• Decision curve and calibration analyses indicated the model's potential clinical utility and accuracy.

Conclusions:

• The developed clinical-radiomics model accurately predicts pCR to NAC in breast cancer.
• This multimodal approach outperforms traditional unimodal methods.
• A dynamic nomogram based on this model can assist in personalizing treatment strategies for BC patients.