Automated image-based detection and grading of lymphocytic infiltration in breast cancer histopathology.

摘要

The identification of phenotypic changes in breast cancer (BC) histopathology is of significant clinical importance in predicting disease outcome and prescribing appropriate therapy. One such example is the presence of lymphocytic infiltration (LI) in histopathology, which has been correlated with a variety of prognoses and theragnoses (i.e. response to treatment) in BC patients. In this thesis work a computer-aided diagnosis (CADx) system is detailed for quantitatively measuring the extent of LI from hematoxylin and eosin (H & E) stained histopathology. The CADx system is subsequently applied to BC patients expressing the HER2 gene (HER2+ BC), where LI extent has been found to correlate with nodal metastasis and distant recurrence. Although LI may be graded qualitatively by BC pathologists, there is currently no quantitative and reproducible method for measuring LI extent in HER2+ BC histopathology. Hence, a CADx system that performs this task will potentially help clinicians predict disease outcome and allow them to make better therapy recommendations for HER2+ BC patients. The CADx methodology comprises three key steps. First, a combination of region-growing and Markov Random Field algorithms is used to detect individual lymphocyte nuclei and isolate areas of LI in digitized H & E stained histopathology images. The centers of individual detected lymphocytes are used as vertices to construct a series of graphs (Voronoi Diagram, Delaunay Triangulation, and Minimum Spanning Tree) and a total of 50 architectural features describing the spatial arrangement of lymphocytes are extracted from each image. By using Graph Embedding, a non-linear dimensionality reduction method, to project the high-dimensional feature vectors into a reduced 3D embedding space, it is possible to visualize the underlying manifold that represents the continuous nature of the LI phenotype. Over a set of 100 randomized cross-validation trials, a Support Vector Machine classifier shows that the architectural feature set distinguishes HER2+ BC histopathology samples containing high and low levels of LI with a classification accuracy greater than 90%.