The landscape ecology of DCIS: Microenvironmental drivers of recurrence and progression to breast cancer.

Abstract

e12568 Background: Modern preventive screening have made it possible to identify ductal carcinoma in situ (DCIS). Progression of DCIS to invasive disease is known to be driven by somatic evolution and affected by clinical interventions. Yet studies indicate that only 20–30% of the cases of DCIS will progress further if not treated, implying that a large portion of patients are over treated. Recognizing which pre-cancerous tumors are likely to progress to cancer or recur as DCIS after initial treatment is of great importance to optimize treatment options, lower associated risks and allocate healthcare efforts to patients that most benefit from early treatment. Methods: This study consists of a retrospective cohort of 158 patients, 57 of which had a DCIS recurrence over 2 to 228 months after treatment, and 45 progressed to invasive disease over 12 to 180 months. We generated a digital pathology library consisting of hematoxylin and eosin (H&E) stained whole slide images (WSI), and applied machine learning methods to perform duct and nuclear segmentation followed by cell classification. We then computed a set of spatial statistics from landscape ecology to characterize each sample. We used LASSO on half of the data (the training set) to select a Cox proportional hazards model that predicted recurrence of DCIS or cancer and used the other half of the data (the test set) to validate the predictors in the Cox model. Results: When considering the time to recurrence as disease-free survival (DFS), the type of DCIS treatment stratifies patients between Lumpectomy Only and More than Lumpectomy groups (log-rank test p-value < 0.0001), recapitulating previous results. When the subset of patients that had more than a lumpectomy was analyzed using spatial landscape ecology metrics, a decrease in the interaction between tumor cells and fibroblasts and between tumor cells and lymphocytes, measured by the mean nearest neighbor distance, were associated with improved DFS (concordance = 0.717) in the model validation set. Stratification of samples by the median risk is discriminatory with a p-value < 0.001. Alternatively, decreased nearest neighbor distance between lymphocytes and fibroblasts was associated with the time to invasive disease (p-value < 0.05). Conclusions: Integrating spatial ecological statistics of cell distributions enhances risk stratification of DCIS samples beyond the known effects of early treatment. Our results highlight the importance of ecological mechanisms in recurrence, progression and treatment resistance. The risk associated with spatial interactions of lymphocytes and fibroblasts suggests previously unknown ecological interactions may be associated with early micro-metastases that result in breast cancer even after mastectomy or lumpectomy with radiation in patients with DCIS.