Automatic Segmentation of Transient Fetal Compartments on Brain Histology

Abstract

Transient fetal compartments that appear during prenatal brain development are of utmost importance for cortical layers and white matter formation. However, manual delineation of these structures is expensive and time-consuming, resulting in a lack of readily available datasets for training supervised machine learning segmentation models. Here, we present a pipeline for coarse, self-supervised segmentation of these compartments in 2D Nissl-stained histological slices of the human fetal brain. Preprocessing involves individual cell segmentation and the creation of a cell-graph. Then, a custom self-supervised graph neural network which is utilizing a combined contrastive and rank-ordering loss generates cell embeddings that encode compartment-wise differences. Finally, a Gaussian mixture model clusters these embeddings into discrete groups providing initial segmentations that can be efficiently corrected by experts, thereby significantly speeding up the manual delineation process. Our method shows promising results with average Dice scores, calculated between the models output and experts correction of that output, of 72.89%, 80.83%, 91.30% and 78.90% for 13, 15, 17 and 21 postconception weeks, respectively.