Breast-conserving surgery (lumpectomy) has become the standard local treatment for ductal carcinoma in situ (DCIS), offering patients the chance to preserve their breast while removing the tumor with clear margins. However, a meaningful proportion of patients experience disease recurrence after this procedure, including both recurrent DCIS and progression to invasive breast cancer.
Current clinical risk tools such as tumor size, surgical margin width, age, and histological grade have limited ability to precisely identify which post-lumpectomy DCIS patients will develop long-term recurrence. More advanced molecular tools exist, like the DCIS Score gene expression assay, but these are costly and not universally available.
Mammography is already the standard imaging modality for DCIS detection, and over 90% of DCIS cases, especially high-grade disease, show suspicious microcalcifications on mammograms. The systematic visual information captured in these images represents an underutilized source of prognostic data. Deep learning can extract complex quantitative patterns from mammograms, known as radiomic features, that may correlate with biological aggressiveness and recurrence risk.
A key challenge with complex machine learning models is interpretability. Clinicians need to understand not just what a model predicts, but why, in order to trust and act on the predictions. SHAP (Shapley Additive Explanations) analysis provides a rigorous framework for attributing each model prediction to specific input features, transforming opaque AI outputs into transparent, clinician-friendly explanations.
The study retrospectively analyzed 140 female DCIS patients treated at Harbin Medical University Cancer Hospital between 2011 and 2020, with follow-up extending to April 2025. All patients had biopsy-confirmed DCIS, underwent breast-conserving surgery with pathologically negative margins, and had complete digital mammography images and clinical data available.
An extensive set of clinical variables was collected including patient demographics, tumor characteristics (histological grade, Ki-67 proliferation index, hormone receptor and HER2 status), treatment details (chemotherapy, radiotherapy, endocrine therapy), and four serum inflammatory and immunity markers: platelet-to-lymphocyte ratio (PLR), neutrophil-to-lymphocyte ratio (NLR), lymphocyte-to-monocyte ratio (LMR), and platelet-albumin ratio (PAR). These immune biomarkers capture systemic inflammatory state at diagnosis, which may influence tumor biology and recurrence risk.
Digital mammograms from two types of imaging systems were collected, with radiologists independently evaluating each image using the standardized BI-RADS (Breast Imaging Reporting and Data System) classification. Calcified regions were manually annotated by two trained researchers supervised by a senior breast radiologist using 3D Slicer software. The annotated regions of interest served as the input areas for radiomic feature extraction.
Recurrence was defined broadly to include any ipsilateral breast recurrence, chest wall involvement, regional lymph node involvement, or distant metastasis. Patients were followed at 3-month intervals for the first 6 months, then 6-monthly for up to 5 years, then annually. The dataset was split 10:4 into 100 training cases and 40 validation cases.
Radiomic feature extraction from mammography images was performed using the pyradiomics Python package, generating an initial pool of 849 quantitative imaging features including shape descriptors, first-order histogram statistics, and second-order texture features such as those derived from the gray-level co-occurrence matrix. These features capture mathematical properties of the image intensity distribution within the annotated calcification region.
A rigorous multi-step feature selection pipeline reduced this to five final mammographic features. First, Spearman correlation analysis removed features correlated above 0.9 with other features, reducing redundancy to 172 features. Mann-Whitney U testing then identified 41 features significantly different between recurrent and non-recurrent groups. Finally, LASSO regression with 10-fold cross-validation selected the five most predictive mammographic features for the mammographic signature (MS).
The five selected features were: 10th percentile pixel intensity, gray-level contrast, inverse difference moment normalized, and two wavelet-filtered image median values. These features collectively capture the intensity distribution, textural homogeneity, and multi-scale structural properties of the calcification pattern, providing a composite imaging fingerprint of the lesion.
The mammographic signature was then combined with clinical variables and the four immune biomarkers. Three independent feature selection methods (SVM-RFE, LASSO, and random forest) were applied to this expanded feature set. The four variables appearing consistently across all three methods were identified as the core predictors: mammographic signature, Ki-67 index, histological grade, and chemotherapy status. Sixteen machine learning algorithms were then trained on these four variables.
The Gradient Boosting Machine (GBM) algorithm achieved the best overall performance among all 16 models evaluated. On the training set, it achieved an AUC of 0.963 with sensitivity of 0.958 and specificity of 0.933. On the independent validation test set, it achieved an AUC of 0.918 (95% CI: 0.873-0.963) with sensitivity of 0.900 and specificity of 0.857, demonstrating strong generalization beyond the training data.
Several models showed extreme overfitting, notably Decision Tree and Extra Trees, which achieved perfect training AUC of 1.000 but dropped substantially on test data. This pattern illustrates why model selection must prioritize test set rather than training performance. In contrast, GBM maintained high and consistent performance across both sets, reflecting its built-in regularization through tree depth constraints and gradient-based optimization.
Comparing the GBM integrated model against each individual risk signature alone confirmed the value of combining modalities. The mammographic signature alone achieved training AUC of 0.937 and test AUC around 0.849. Integrating all four variables into GBM improved test AUC to 0.918, confirming that clinical features and imaging features carry complementary rather than redundant information.
Decision Curve Analysis (DCA) validated the model's clinical utility by showing that the GBM model provides positive net benefit across a wide range of probability thresholds for treatment decision-making. This analysis ensures that the model helps clinicians make better decisions than the alternatives of treating everyone or treating no one, across the full spectrum of clinician risk tolerance.
SHAP (Shapley Additive Explanations) analysis ranked the four predictors by their average absolute contribution to model predictions: mammographic signature (MS) ranked first, followed by Ki-67 index, histological grade, and chemotherapy status. This confirms that the imaging-derived features contribute more predictive information than any single clinical variable.
SHAP violin plots revealed how each feature influences predictions directionally. Higher MS values (indicating more complex or suspicious calcification patterns such as fine linear or branching microcalcifications) were associated with increased predicted recurrence risk, while lower MS values (more benign-like features) decreased predicted risk. Elevated Ki-67 and higher histological grade similarly increased predicted recurrence probability.
The value of SHAP at the individual patient level was demonstrated through a force plot analysis of a representative case. For one patient, a high histological grade was the primary factor reducing the predicted recurrence risk below the population baseline (shifting the predicted probability from 0.243 to 0.144), while absence of chemotherapy slightly increased predicted risk. This individual-level transparency transforms the model into a practical clinical communication tool.
Importantly, SHAP analysis identified a high-risk patient profile: high MS combined with elevated Ki-67 above 30% and high histological grade indicates an extremely high-risk patient, even with negative surgical margins. For these patients, standard adjuvant radiotherapy alone may be insufficient, suggesting need for intensified monitoring or additional systemic therapy. Conversely, low MS, low grade, and low Ki-67 together identify candidates for de-escalated treatment.
The study's finding that patients not receiving radiotherapy had markedly higher recurrence rates is consistent with published literature showing adjuvant radiotherapy reduces DCIS recurrence risk by 50 to 70% after lumpectomy. However, SHAP analysis added nuance: even among patients who received radiotherapy, those with high-grade lesions and high Ki-67 remained at elevated risk, suggesting that tumor biology can partially overcome the benefits of radiotherapy.
The combined high-risk profile of high mammographic signature, high Ki-67, and high grade represents patients whose tumor biology is inherently aggressive. For these patients, the model suggests considering additional interventions beyond standard adjuvant radiotherapy, including optimized radiotherapy dosing, extended endocrine therapy, systemic chemotherapy, or intensified surveillance protocols.
The model's advantage over existing clinical scoring systems such as the VNPI (Van Nuys Prognostic Index) is its ability to capture tumor heterogeneity through imaging features that traditional scoring systems based on size, margin, and grade alone cannot quantify. The deep learning-derived mammographic signature is sensitive to the spatial distribution and morphology of calcifications in ways that human visual scoring with fixed descriptors is not.
This study successfully developed an interpretable GBM model achieving AUC 0.918 for predicting 5-year DCIS recurrence after lumpectomy by combining deep learning mammographic features with clinical and laboratory data. The integration of SHAP methodology transforms the otherwise opaque GBM model into a transparent decision-support tool that clinicians can audit and trust.
Key limitations include the small sample size (140 patients) relative to the initial feature dimensionality, which required rigorous dimensionality reduction strategies. The single-institution retrospective design from a Chinese population limits external validity, and results may not generalize to other racial and ethnic groups. Treatment protocols were also not entirely uniform, with incomplete documentation of specific radiotherapy dosing and chemotherapy regimens.
Future validation in multicenter prospective cohorts spanning diverse populations is essential before clinical implementation. Incorporating richer imaging modalities such as dynamic contrast-enhanced MRI could further improve the model's predictive power. The authors propose this model as an auditable decision-support tool, not a replacement for clinical judgment, enabling personalized postoperative management strategies for DCIS patients following breast-conserving surgery.