Quantitative detection and stratification of lymphocytic infiltration in breast cancer
Ajay, Nagesh, Basavanhally ; Rutgers University;
Content:
Molecular changes in breast cancer (BC) are often accompanied by changes in the tumor microenvironment. One such example is the presence of lymphocytic infiltration (LI), a histological feature that has been recently demonstrated to correlate with prognosis in early stage HER2+ breast cancer. The evaluation of LI in BC histology is usually measured in a qualitative manner that can have high intra-observer variation. A robust quantitative scheme would create a more precise stratification and allow for objective analysis of the extent of LI present in a tumor and its relation to clinical outcome measures. In this study, we develop computer algorithms to measure the extent of LI in BC.
Technology:
We utilize 40 images of H&E stained histology obtained from 7 HER2+ breast cancer samples. Each image was examined by an experienced clinician and scored as having low, intermediate or high level of lymphocytic infiltration. All samples are digitized via a whole-slide scanner and analyzed with the Matlab software package.
Design:
We apply an automated segmentation scheme to find the centroid of each lymphocyte. We calculate graph-based features (Voronoi Diagram, Delaunay Triangulation, and Minimum Spanning Tree) from these centroids to quantify the architecture (i.e. relative spatial placement) of lymphocytes. We calculate several feature, including lymphocyte density, area disorder, and nearest neighbor statistics. Manifold learning techniques, such as graph embedding, allow us to visualize the stratification of samples in a low-dimensional feature space. We further apply a classification algorithm, known as support vector machine, to provide statistical results.
Results:
The result of applying manifold learning to the graph-based features from manually-segmented lymphocytes is shown in the figure below. Each point represents a sample, and proximity between samples denotes similarity in reduced feature space. The low-dimensional manifold reveals the underlying structure of the data by showing the progression from low to high degrees of infiltration. Furthermore, randomized classification demonstrates an accuracy of 89.50% ± 6.22% for discriminating between low and high LI samples using manually-segmented graph-based features.
Conclusion:
We have presented an automated algorithm for detecting and classifying LI in high grade BC. We have used graph-based features to exploit the architectural differences in arrangement between BC nuclei and lymphocytes. The graph embedding manifold shows the potential for using graph-based features in determining the degree of LI, rather than simply the presence or absence of LI.
