Deep Learning Classification of Grade 2 Nottingham Breast Tumors via DCE MRI

Overview

A deep learning (DL) model using dynamic contrast-enhanced (DCE) MRI was developed to classify breast tumors by Nottingham Histologic Grade (NHG), focusing on the heterogeneous NHG2 group. The model successfully stratified NHG2 tumors into NHG1-like and NHG3-like subgroups with distinct recurrence-free survival outcomes, providing prognostic value beyond conventional clinical predictors.

Background

The Nottingham Histologic Grade (NHG) is a key prognostic factor in breast cancer, categorizing tumors into NHG1, NHG2, or NHG3 based on histopathological features. NHG2 tumors represent about half of breast cancers and exhibit heterogeneous biological behavior, complicating treatment decisions. Molecular profiling can further stratify NHG2 tumors but is costly and not routinely feasible. Deep learning approaches applied to imaging, particularly DCE MRI, offer a non-invasive, cost-effective method to refine risk stratification and guide personalized therapy.

Data Highlights

Key Findings

• The DL model was trained on NHG1 and NHG3 tumors from the Duke dataset and externally validated on the Advanced-MRI dataset.
• The model classified NHG2 tumors into NHG1-like and NHG3-like subgroups based on DCE MRI features.
• Recurrence-free survival differed significantly between the NHG2 subgroups, indicating prognostic relevance.
• The DL-based classification provided prognostic information independent of age, lymph node status, tumor stage, and molecular subtype.
• Using 2D CNN architecture with cropped tumor regions and data augmentation optimized model performance and minimized overfitting.

Clinical Implications

This DL approach enables non-invasive, cost-effective risk stratification of NHG2 breast tumors using routinely acquired DCE MRI, potentially improving personalized treatment decisions. Patients with NHG3-like tumors may benefit from intensified therapy, while those with NHG1-like tumors could avoid overtreatment and associated toxicities. Integration of this imaging-based tool could complement existing clinical and molecular assessments to refine prognosis.

Conclusion

The study demonstrates that a DL model applied to DCE MRI can effectively subclassify heterogeneous NHG2 breast tumors into prognostically distinct groups, enhancing risk stratification beyond traditional clinical factors. This approach holds promise for guiding more tailored treatment strategies in breast cancer care.

References

1. Nottingham Histologic Grade and Breast Cancer Prognosis
2. Duke-Breast-Cancer-MRI Dataset
3. Advanced-MRI-Breast-Lesions Dataset