Deep Learning-Based Subtyping of Early-Stage Breast Cancer from Mammography: Distinguishing Luminal and Non-Luminal Tumors for Precision Therapy
Abstract
Purpose
Breast cancer is the second leading cause of cancer death among women. Fortunately, for this biologically heterogeneous disease, several molecular subtypes corresponding to distinct responses to treatments and prognoses have been recognized, enabling subtype-directed precision therapy. Compared with luminal subtypes, HER2-enriched and triple-negative subtypes are more sensitive to neoadjuvant treatment. Thus, neoadjuvant therapy before surgery is recommended for clinical T1 and T2 non-luminal breast cancers to acquire the treatment response and pathological complete response (pCR) status. In this study, we developed and validated a deep learning model using pretreatment mammography to distinguish between luminal and non-luminal breast cancers at early stages with the ultimate goal of supporting timely clinical decision-making on the use of neoadjuvant therapy.
Methods
Between February 2015 and December 2021, a total of 848 mammograms from 424 patients with pathology-confirmed early-stage breast cancer were retrospectively collected at our institution. Both CC and MLO views of the breast were included. The ground truth was labelled on the basis of breast surgical pathology-confirmed results. A Convnext-small network was employed as the backbone to extract view-specific image features and pretrained weights from a larger domain-specific dataset were first used for initialization, followed by further training using our mammography dataset. The subtyping performance was evaluated at the breast level, with the breast-level prediction defined as the average of the predicted probabilities from the two views.
Results
The model demonstrated robust performance. Through four-fold cross-validation, the model achieved an average area under the curve (AUC) of 0.758 across the internal validation cohorts. In the independent internal test cohort, the model yielded an AUC of 0.711 (95% CI, 0.630–0.787).
Conclusion
Our model can effectively distinguish between luminal and non-luminal breast cancers at early stages before surgery based on mammogram, providing a tool to facilitate individualized treatment decision making.