Figure 1. Instance segmentation of immune cells in high-resolution fluorescence microscopy images of LN kidney biopsies.

(A) Automatic instance segmentation of 5 immune cell classes was performed by combining predictions from 2 instances of Mask R-CNN: one trained to segment CD20⁺, CD3⁺CD4^–, and CD3⁺CD4⁺ lymphocytes and one trained to segment pDCs and mDCs. Cell location, class, and morphological features were calculated from joint predictions. (B) The Mask R-CNN architecture comprises a ResNet Feature Pyramid Network (FPN) backbone used for feature extraction, a region proposal network (RPN) used to generate cell proposals, and two parallel branches used for (a) semantic segmentation (mask branch) and (b) classification (softmax layer) and localization (bounding box [Bbox] regression) of cell proposals. (C) Representative segmentations produced by the multinetwork pipeline showed strong agreement with the expert-defined manual segmentations. Magnification for all images 63x (zoom factor 1.7). This figure was created with BioRender.com.