Deep Learning Radiomics Does Not Improve Residual Cancer Burden Prediction Post-NAC in Breast MRI

Overview

This study evaluated deep learning-based radiomics for predicting residual cancer burden (RCB) after neoadjuvant chemotherapy (NAC) in locally advanced breast cancer using breast MRI. Results showed that deep radiomics did not outperform conventional clinical predictors such as tumor volume and subtype in assessing RCB prior to surgery.

Background

Neoadjuvant chemotherapy is increasingly used for locally advanced breast cancer, with pathological complete response (pCR) and residual cancer burden (RCB) being important prognostic markers. Reliable imaging predictors of tumor response to NAC remain elusive despite the sensitivity of dynamic contrast-enhanced MRI (DCE-MRI). Machine learning approaches, including radiomics and deep learning, have been explored to improve prediction but have yet to demonstrate clear clinical benefit. This study aimed to assess whether deep radiomics could enhance RCB prediction beyond standard clinical parameters.

Data Highlights

Key Findings

• Deep learning-based radiomics features extracted automatically from DCE-MRI did not improve prediction of residual cancer burden compared to conventional clinical parameters.
• Tumor volume and receptor subtype remained the strongest predictors of RCB after NAC.
• Neither deep radiomics nor traditional imaging biomarkers provided sufficient specificity to reliably identify pathological complete response.
• Multi-institutional validation confirmed the limited added value of deep radiomics in this clinical setting.
• Sampling errors and heterogeneity in imaging acquisition protocols may contribute to challenges in predictive modeling.

Clinical Implications

Clinicians should continue to rely on established clinical and pathological parameters such as tumor volume and receptor subtype when assessing residual disease after NAC. Current deep learning radiomics approaches do not yet provide sufficient accuracy to guide surgical decision-making or safely omit surgery. Further research is needed to improve imaging biomarkers and integrate multimodal data for better prediction of treatment response.

Conclusion

Deep learning radiomics applied to breast DCE-MRI does not enhance prediction of residual cancer burden following neoadjuvant chemotherapy beyond standard clinical predictors. Conventional parameters remain essential for guiding clinical management in locally advanced breast cancer.

References

1. Symmans et al. -- Residual Cancer Burden Scoring
2. LIMA Study Protocol -- Liquid biopsies and Imaging in Breast Cancer
3. Breast Imaging Reporting and Data System (BI-RADS)