DeepSlice: a deep neural network for rapid fully automatic registration of histological images to the Allen Mouse Brain Atlas
The Allen Mouse Brain Atlas Common Coordinate Framework (CCF) provides a standardised frame of reference to which gene expression, connectivity, and functional data from experiments conducted in mice are routinely registered by neuroscientists. At present, registration to the CCF is a labour-intensive process, in which the aligner typically compares their histological images to a volumetric brain atlas, manually aligning cutting angles, scale, rotation and rostrocaudal position until atlas and histology match. This process is demanding of time and expertise, taking months of training to gain proficiency and, even then, several minutes to align each section (i.e. several hours per mouse brain). To solve this bottleneck, we have trained a deep convolutional neural network (CNN), DeepSlice, to predict the CCF coordinates of a diverse set of mouse brain images without need of preprocessing or operator input. We first developed a curated dataset of mouse brain histology to quantify human registration performance across multiple levels of expertise, which we could use as a benchmark for assessing the algorithm. We then trained a CNN to register histological images acquired from the Allen Gene Expression Library, and synthetic sections which we had generated. We show that CNNs are capable of rapidly registering brain histology at expert level. DeepSlice differs from previous attempts to automate registration both in terms of its simplicity (no user training required) and its speed (milliseconds per section on a consumer-level laptop). DeepSlice, now running as a live, fully-functional web application, provides a tool that will allow scientists to painlessly register their work or reanalyse enormous public datasets, enhancing efficiency, transparency and replicability.