A deep learning framework to stratify Nottingham histologic grade 2 breast tumors based on dynamic contrast-enhanced MRI.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, Montefiore Health System and Albert Einstein College of Medicine, Bronx, NY, USA.
- Department of Oncology, Montefiore Health System and Albert Einstein College of Medicine, Bronx, NY, USA.
- Department of Pathology, Montefiore Health System and Albert Einstein College of Medicine, Bronx, NY, USA.
- Department of Radiology, Montefiore Health System and Albert Einstein College of Medicine, Bronx, NY, USA. [email protected].
Abstract
The Nottingham Histologic Grade (NHG) informs prognosis and treatment decisions in breast cancer, but NHG2 tumors are biologically heterogeneous, leading to both under- and over-treatment. Clinical and imaging data from the Duke-Breast-Cancer-MRI (n = 877) and advanced-MRI-breast-lesions (n = 37) datasets were used to develop DeepRadGrade (DRG), a convolutional neural network model trained to differentiate NHG1 from NHG3 tumors on dynamic contrast-enhanced (DCE) MRI. The model then classified 456 NHG2 tumors into DRG2- (NHG1-like) and DRG2+ (NHG3-like) subgroups. Recurrence-free survival (RFS) was assessed with Kaplan-Meier and Cox models adjusting for age, lymph node invasion, tumor stage, and molecular subtype. DRG achieved an AUC of 0.84 [95% CI: 0.83-0.86] in training, 0.82 [0.71-0.91] in testing, and 0.84 [0.69-0.96] in external validation. Among NHG2 tumors, 315 were classified as DRG2- and 131 as DRG2+. Patients with DRG2+ tumors had significantly worse RFS compared to DRG2- (adjusted hazard ratio = 2.39 [95% CI: 1.29-4.45], p = 0.0059), independent of standard prognostic factors. Incorporating DRG classification improved the Cox model's C-index from 0.68 to 0.73 (p = 0.040). A deep learning model applied to routine DCE MRI effectively stratified NHG2 breast tumors into clinically meaningful subgroups with distinct recurrence risk. This approach offers a cost-effective tool for individualized risk stratification and could help tailor treatment to minimize over- and under-treatment in intermediate-grade breast cancer. Question Difficulty in deciding between treatment options (neoadjuvant chemotherapy or primary surgery) in patients with intermediate risk breast cancer (Nottingham Grade 2). Findings Deep learning model reclassified grade 2 tumors into grade 1- and 3-like based on MRI. Patients with grade 3-like tumors had worse RFS: adjusted hazard ratio = 2.39 [95% CI: 1.29-4.45], p = 0.0059). Clinical relevance Risk stratification could inform treatment choice to minimize over- and under-treatment in patients with intermediate-risk breast cancer.