Deepslice: a deep neural network for fully automatic alignment of histological sections

An almost ubiquitous component of laboratory neuroscience research is mapping the locations of neurons based on their function, genetic profile or connectivity. An obvious but fundamental aspect of this task is accurately documenting the locations of neurons of interest. In thousands of laboratories across the world, this is done in a subjective and time-consuming process in which histological sections are compared to 2- or 3-dimensional reference atlases, such as the Allen Mouse Brain Atlas, and sections are ‘aligned’ to the region of the atlas that most closely resembles their tissue. This task is slow and error-prone, even for experts, with accurate alignment taking several minutes per section. Here we present DeepSlice, a convolutional neural network developed to register histological images of the mouse brain to the Allen Common Coordinate Framework. DeepSlice analyzes histological sections obtained using diverse imaging modalities (block-face confocal, fluorescent immunohistochemistry, NISSL, bright-field in situ hybridization) and aligns them to the Allen Brain Atlas with accuracy comparable to human experts but orders of magnitude faster. Our system requires no user input, can identify imperfections in cutting angle or image rotation, and is computationally efficient, requiring no specialized hardware. Recent technological innovations have revolutionalised the potential scope of neuroscience research, but the process of orienting yourself in the brain remains essentially unchanged for at least 30 years: DeepSlice represents a one-button image alignment tool that requires no specialist expertise and can process thousands of images per minute on a middle-of-the-range laptop, and may therefore unblock this longstanding bottleneck.