CAREER: Scaling up Brain Circuit Reconstruction with Human-centric Machine Learning

The field of connectomics aims to reconstruct the connections between various parts of the brain from extremely high resolution microscopy images. Such a transformative approach can provide detailed renderings of the brain at the cellular level to reveal the organizing principle and the mechanism of neural connectivities. Furthermore, these new insights could accelerate the treatment development for neurodegenerative diseases and inspire novel AI algorithms. However, the connectomics image data of a mere one-millimeter cube brain region is on the petabyte scale, where existing computational pipelines produce too many errors for domain experts to correct in neuron reconstruction. What is missing is not just a better reconstruction method but a human-centric approach to automate the labor-intensive workflows before and after the reconstruction, e.g., data annotation to train the model and error correction to refine the results. This project will build a scalable human-centric computational pipeline with novel algorithms to mimic human cognition to reduce human effort in the pipeline significantly. If successful, the developed workflows will be deployed to expedite the BRAIN Initiative’s ambitious whole-mouse brain reconstruction project to revolutionize the understanding of the brain. This project will focus on accelerating the labor-intensive workflows of data annotation, proofreading, and transfer learning in the machine learning pipeline. Inspired by human cognitive abilities, this project will develop novel machine learning algorithms to exploit various data sources beyond the traditional densely annotated 3D neuron reconstruction. Concretely, this project has the following aims. (1) This project will distill the unlabeled data to learn to group images by appearance to assist domain experts in effectively discovering sub-volumes for annotation and propagating sparse labels to dense reconstruction. (2) This project will build automatic agents to learn from domain experts’ proofreading strategies to detect and correct the automatic reconstruction results. (3) This project will develop transfer learning methods to reuse labeled connectomics datasets and pre-trained models to assist biology labs in analyzing their microscopy images. These three research aims will be accompanied by comprehensive evaluations on collected benchmark datasets and accessible software resources for the biomedical image analysis community. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.