Machine Learning and Immunohistochemistry

The goal of this study was to count the positively stained cell density within rat striatum. Striata were first hand-annotated (Figure 1. in red) to exclude false positives outside the striatum. Then, within the striatum, positively stained regions (Figure 1. in green) were automatically segmented. To do this, a library of 843 reference cells was built to develop a cell classifier using Invicro’s and Sponsor’s expert identification of positively stained cells. A corresponding library of 942 background patches that contain no cells was used as a negative reference for cell classification. A Support Vector Machine Classifier was run for each pixel based on this library. The resulting probability maps were post processed to segment and count the positively stained cells. The resulting counts, along with cell counts from 3 different experts, can be seen below (Figure 2).

Figure 1

Figure 2