Supervised Learning With Perceptual Similarity for Multimodal Gene Expression Registration of a Mouse Brain Atlas

Abstract

The acquisition of high quality maps of gene expression in the rodent brain is of fundamental importance to the neuroscience community. The generation of such datasets relies on registering individual gene expression images to a reference volume, a task encumbered by the diversity of staining techniques employed, and by deformations and artifacts in the soft tissue. Recently, deep learning models have garnered particular interest as a viable alternative to traditional intensity-based algorithms for image registration. In this work, we propose a supervised learning model for general multimodal 2D registration tasks, trained with a perceptual similarity loss on a dataset labeled by a human expert and augmented by synthetic local deformations. We demonstrate the results of our approach on the Allen Mouse Brain Atlas (AMBA), comprising whole brain Nissl and gene expression stains. We show that our framework and design of the loss function result in accurate and smooth predictions. Our model is able to generalize to unseen gene expressions and coronal sections, outperforming traditional intensity-based approaches in aligning complex brain structures.

A method to estimate the cellular composition of the mouse brain from heterogeneous datasets

“…In this study, we present a substantially improved workflow to produce the second version of the Blue Brain Mouse Cell Atlas workflow (BBCAv2). The new workflow replaces the manual alignment of ISH datasets with the image alignment algorithm developed by Krepl et al [30].The pipeline also includes a new method to estimate cell densities from a large body of literature data. In total, we integrated values from 56 different literature sources.…”

“…• For each dataset, every coronal section is automatically realigned to the corresponding anatomical section of the Nissl-stained mouse brain using the Krepl algorithm [30].…”

“…This step was not only labor intensive but also error prone. In our updated workflow, we replaced the manual realignment by a novel deep learning based alignment algorithm by Krepl et al [30].…”

“…For our new pipeline, we therefore replace this step with a fully automatic machine learning based method. The details of this automatic alignment method are described in a separate paper [31]. In brief, this algorithm constructs a displacement field by finding pixel by pixel correspondences between an image and a reference.…”

“…Furthermore, ISH experiments performed with a sagittal slicing covers regions not included in the coronal experiments and their addition will improve the ISH whole brain coverage. Using these datasets requires the realignment of the sagittal images of the inhibitory markers to the Nissl volume, using the deep learning algorithm described from Krepl et al [31]. The usage of these datasets is currently investigated for a future update of the pipeline.…”