Deep Learning Based Automatic Immune Cell Detection for Immunohistochemistry Images

Abstract

Immunohistochemistry (IHC) staining is a widely used technique in the diagnosis of abnormal cells such as cancer. For instance, it can be used to determine the distribution and localization of the differentially expressed biomarkers of immune cells (such as T-cells or B-cells) in cancerous tissue for an immune response study. Typically, the immunological data of interest includes the type, density and location of the immune cells within the tumor samples; this data is of particular interest to pathologists for accurate patient survival prediction. However, to manually count each subset of immune cells under a bright-field microscope for each piece of IHC stained tissue is usually extremely tedious and time consuming. This makes automatic detection very attractive, but it can be very challenging due to the wide variety of cell appearances resulting from different tissue types, block cuttings, and staining processes. This paper presents a novel method for automatic immune cell counting on digitally scanned images of IHC stained slides. The method first uses a sparse color unmixing technique to separate the IHC image into multiple color channels that correspond to different cell structures. Since the immune cell biomarkers that we are interested in are membrane markers, the detection problem is formulated into a deep learning framework using the membrane image channel. The algorithm is evaluated on a clinical data set containing a large number of IHC slides and demonstrates more effective detection than the existing technique and the result is also in accordance with the human observer’s output.