Identifying Partial Mouse Brain Microscopy Images from the Allen Reference Atlas Using a Contrastively Learned Semantic Space

Registering mouse brain microscopy images to a reference atlas is crucial to determine the locations of anatomical structures in the brain, which is an essential step for understanding the function of brain circuits. Most existing registration pipelines assume the identity of the reference plate – to which the image slice is to be registered – is known beforehand. This might not always be the case due to three main challenges in microscopy image data: missing image regions (partial data), different cutting angles compared to the atlas plates and a large number of high-resolution images to be identified. Manual identification of reference plates as an initial step requires highly experienced personnel and can be biased, tedious and resource intensive. On the other hand, registering images to all atlas plates can be slow, limiting the application of automated registration methods when dealing with high-resolution image data. This work proposes to perform the image identification by learning a low-dimensional space that captures the similarity between microscopy images and the reference atlas plates. We employ Convolutional Neural Networks (CNNs), in the Siamese network configuration, to first obtain low-dimensional embeddings of microscopy image data and atlas plates. These embeddings are contrasted with positive and negative examples in order to learn a semantically meaningful space that can be used for identifying corresponding 2D atlas plates. At inference, atlas plates that are closest to the microscopy image data in the learned embedding space are presented as candidates for registration. Our method achieved TOP-3 and TOP-5 accuracy of 83.3% and 100%, respectively, compared to the SimpleElastix-based baseline which obtained 25% in both the Top-3 and Top-5 accuracy (Source code is available at https://github.com/Justinas256/2d-mouse-brain-identification ).