Ductal carcinoma in situ (DCIS) accounts for approximately 20% of screen-detected breast cancers, and the major risk after treatment stems almost entirely from local recurrence in the same breast. After breast-conserving surgery (lumpectomy), patients face a meaningful risk of the cancer returning, either as DCIS again or as the more dangerous invasive form.
Standard clinicopathological variables used to assess recurrence risk, such as tumor size, margin status, nuclear grade, presence of comedo necrosis, and patient age, have demonstrated only limited ability to reliably identify which patients will recur. Even combinations of these factors, such as the Van Nuys Prognostic Index and various prognostic nomograms, have not provided sufficiently accurate risk stratification to guide treatment decisions.
Molecular tools such as gene expression assays have been explored but have not shown consistent results, lack the ability to account for the tumor microenvironment (the surrounding stromal tissue, immune cells, and blood vessels), and are costly. Without a reliable risk tool, clinicians cannot confidently distinguish which patients truly need adjuvant radiotherapy and which patients could safely avoid it, leading to both overtreatment and undertreatment.
Computational pathology, which extracts quantitative features from digitized tissue slides, offers a promising path forward. Whole slide image analysis can capture the full spatial architecture of a tumor section, including information from stromal regions, immune infiltrates, and vascular patterns that standard pathological grading ignores. This study developed a two-stage machine learning pipeline using whole slide images to predict 10-year DCIS recurrence risk.
The study used a retrospective cohort of 344 DCIS patients treated with breast-conserving surgery at Nottingham City Hospital between 1989 and 2012, with follow-up data for 10-year recurrence outcomes. All patients had pure DCIS without any invasive component, were treated with lumpectomy rather than mastectomy, and had complete digitized slides available. The cohort was split into a training set of 159 patients (31 with recurrence) and an independent validation set of 185 patients (26 with recurrence).
Formalin-fixed paraffin-embedded tumor blocks were cut into 4-micrometer sections, stained with hematoxylin and eosin (H&E), the standard clinical staining that colors cell nuclei blue-purple and surrounding tissue pink. Sections were scanned at 40x magnification, producing whole slide digital images of approximately 0.24 micrometers per pixel. Slides with quality issues such as out-of-focus areas were rescanned, and those with folded tissue were excluded.
The pipeline operated in two sequential stages. In the first stage, a random forest classifier trained by pathologists automatically annotated each slide into five tissue classes: stroma, normal or benign epithelial ducts, cancerous DCIS ducts, dense lymphocyte regions (defined as regions where immune cells occupied at least 50% of the tile area), and blood vessels. Areas not belonging to any of these classes, such as fat, were labeled as background.
For the annotation classifier, pathologists manually selected ground truth image tiles of 50 x 50 pixels from 32 training slides, yielding 10,359 tiles that were augmented fourfold through rotation and flipping to 41,436 tiles. From each tile, 166 texture features were extracted from the hematoxylin channel after color deconvolution, which mathematically separates the blue nuclear stain from the pink eosin stain to isolate nuclear and chromatin information.
Following automated slide annotation, a comprehensive set of full-slide features was extracted from each annotated tissue class. The majority of these features (99%) captured statistical distributions of the 166 texture features within each annotated class, using four statistical moments: mean, standard deviation, skewness, and kurtosis. Additional spatial features described the distance and size relationships between cancer regions and either blood vessels or immune-dense stromal areas, which prior research suggested had prognostic relevance.
Setting a 10-year follow-up threshold for recurrence status maximized the number of features that differed significantly between patients who recurred and those who did not. At this threshold, approximately 1,238 features (37% of all whole-slide features) showed significant differences between groups. A sequential forward feature selection algorithm then identified the smallest subset of features that achieved the best classification performance, yielding a final model of eight optimized features.
Of the eight selected features, half originated from cancer duct annotations and half from surrounding tissue regions including stroma and benign ducts. This balance highlights that both the tumor itself and its microenvironment carry independent prognostic information. The most important single feature quantified the average hematoxylin (nuclear stain) intensity per pixel within malignant ductal areas above an automatically determined threshold, effectively capturing whether cancer ducts had an open lumen (less dense staining) versus a more solid architecture (denser staining).
The recurrence risk classifier was a random forest model trained on these eight features with equal class probabilities to avoid bias from the class imbalance between recurrent and non-recurrent patients. Performance was assessed through 100 iterations of five-fold cross-validation on the training cohort, and then applied without retraining to the independent validation cohort. The annotation step proved critical: a model built without the prior tissue annotation step achieved a substantially lower hazard ratio (2.82 versus 3.19), confirming that classifying tissue context improves recurrence prediction.
On the training cohort, the eight-feature model achieved an accuracy of 0.87, sensitivity of 0.71, specificity of 0.91, and a hazard ratio of 11.6 (95% CI: 5.3-25.3), meaning patients classified as high-risk were over eleven times more likely to experience recurrence within 10 years than low-risk patients. The high-risk group had only a 35% recurrence-free survival at 10 years, compared to 93% in the low-risk group, a dramatic and clinically meaningful separation.
On the independent validation cohort, the model maintained strong performance: accuracy 0.85, sensitivity 0.50, specificity 0.91, and a hazard ratio of 6.4 (95% CI: 3.0-13.8, p less than 0.0001). Ninety-two percent of patients classified as low-risk remained recurrence-free for 10 years (mean recurrence-free time of 112 months), while only 54% of high-risk patients stayed recurrence-free (mean 73 months). The model retained significance after multivariate adjustment for necrosis, tumor size, margin status, and patient age.
Compared to standard clinical variables, the eight-feature model provided superior accuracy, specificity, positive predictive value (PPV of 0.46 vs. 0.05-0.15 for individual clinical variables), negative predictive value, and odds ratio in the validation cohort. Individual clinical variables all showed substantially lower accuracy and PPV despite sometimes having higher sensitivity, meaning the model was far better at correctly identifying which patients in the high-risk group would actually recur rather than just flagging a large proportion of the population as high-risk.
The model also significantly stratified both high-grade and low/intermediate-grade DCIS patients into distinct risk groups, as well as patients treated with BCS alone and those who additionally received adjuvant radiotherapy. Even among irradiated patients, a high-risk subgroup was identifiable, suggesting that some patients' tumor biology confers elevated risk that radiotherapy does not fully overcome.
The most prognostically important feature in the model quantifies hematoxylin intensity distribution within cancer duct regions at a high-end threshold, which effectively distinguishes DCIS with open lumens (such as cribriform architecture) from DCIS with more solid growth patterns. Lower values of this feature, indicating more solid, densely packed ductal profiles, were strongly associated with higher recurrence risk, consistent with established pathology showing that solid DCIS patterns are more aggressive and often higher grade.
The presence of features derived from surrounding stroma, benign epithelial ducts, and blood vessel regions in the final model demonstrates that the tumor microenvironment contributes independent prognostic information beyond the cancer cells themselves. Stromal cells including fibroblasts and myofibroblasts show distinct hematoxylin staining patterns reflecting their different cellular compositions, and both cell types have been previously implicated in DCIS invasion and recurrence.
Notably, immune-dense regions (dense lymphocyte infiltrates) did not appear among the significant features, likely because lymphocytes show relatively uniform nuclear morphology with limited variability in hematoxylin texture between patients. This contrasts with cancer and stromal regions, which show substantial heterogeneity that carries prognostic signal. The feature that captured benign epithelial duct texture may additionally reflect premalignant columnar cell lesions and usual ductal hyperplasia that co-occur with DCIS and represent additional markers of cancer progression risk.
The model's ability to identify a low-risk DCIS group with only an 8% 10-year recurrence risk even without radiation compares favorably to the Oncotype DX DCIS Score, the currently available commercial tool for DCIS risk stratification, which identifies a low-risk group with a 10.6% 10-year recurrence rate. The model also produced hazard ratios more than double those reported with the Oncotype DX DCIS Score, suggesting stronger risk separation capability.
For high-risk patients identified by the model, particularly those in whom high-risk classification persists despite adjuvant radiotherapy, the model suggests a need for more aggressive interventions, such as intensified surveillance, systemic therapy, or more extensive surgical approaches. For patients flagged as low-risk, the model provides a basis for safely avoiding adjuvant radiation, reducing treatment-related side effects without compromising outcomes.
A key advantage over existing molecular and genomic tools is that the model works entirely from a standard H&E-stained tissue section, which is already prepared for every DCIS patient as part of routine pathology workflow. No additional staining, molecular testing, or special equipment is required beyond a digital slide scanner, making the approach potentially far more accessible and cost-effective than gene expression assays.
This study demonstrates that a two-stage machine learning pipeline using whole slide image analysis can predict 10-year DCIS recurrence risk significantly better than any individual clinical variable, achieving hazard ratios of 11.6 and 6.4 in training and independent validation cohorts respectively. The pipeline's ability to capture spatial relationships between cancer cells, stroma, immune infiltrates, and blood vessels represents a fundamentally new type of prognostic information.
Key limitations include that both cohorts originate from the same institution, limiting evidence of generalizability across different tissue processing protocols and imaging systems. The validation cohort comprised exclusively high-grade DCIS patients, leaving the model's performance in lower-grade disease supported primarily by training cohort data. The training and validation cohorts also differed significantly in comedo necrosis rates and grade distributions, which may have reduced model performance in the validation set relative to training.
Future directions include validation in multi-institutional cohorts with diverse grade distributions and treatment protocols, retraining the model on core biopsy rather than surgical excision specimens to enable pre-treatment risk stratification, and exploring deep learning convolutional network approaches that may identify morphological signatures beyond what hand-engineered texture features can capture. The authors propose this pipeline as a cost-effective clinical tool that could directly inform treatment decisions including the recommendation or omission of adjuvant radiotherapy.