Abstract:
Cancer heterogeneity is associated with drug resistance, risk of disease recurrence and metastasis. Pathology allows for the assessment of heterogeneity in the tumor. However, this assessment is limited to an overall assessment of the tumor microenvironment (TME), through measurement of mitotic levels and cancer invasion. We present PhenoScope, a platform for multi-scale analysis and visualization of tumor heterogeneity that integrates cancer data across scales to extract cross modality trends that drive cancer invasion.
Here we present a Phenoscope analysis of breast tumor pathology slides at three scales. We developed and validated 3 convolutional neural networks (CNN). We combined the outputs of these networks with 2D simulations of the metabolic behavior and growth of cells within the TME. Then, we developed 2 CNNs and implemented an existing CNN that 1) identify cells undergoing mitosis, 2) segment individual cells and classifies their type, and 3) segment five tissues from pathology slides. Additionally, we used transcriptional data to generate patient specific metabolic models which were used with CNN outputs to predict spatial interactions within the TME. Our results showed:
- Mitosis detection CNN: accuracy of 76.2% (precision=83%, recall=76%) in the test set.
- Classification CNN: Dice Similarity Coefficient (DSC) of 0.821 for segmenting cells and an F1 score for classifying cells ranging from 0.559 (F1i) to 0.756 (F1d).
- Segmentation CNN: accuracy of 78.1% with DSC ranging from 0.66 and 0.86 depending on tissue.
The segmentations were analyzed by a proprietary simulation framework along with patient-specific metabolic models to predict the spatial gradients of nutrients, and spatial organization of growth and metabolism. Our simulations identified different metabolic regions in breast tumors dependent on lactate production or consumption, and alanine uptake. These regions were correlated with other indicators such as mitoses and immune cell infiltration. Overall, we demonstrate a proof-of-principle approach of integrating data across scales, allowing for novel predictions and insights of TME behavior in breast cancer.
View the full publication here.
