Design Day 2023 Project List

Proofreader Evaluation and Analysis toolKit (PEAK) for Nanoscale Connectomics

Team: JHU APL: AGT Cohort

Course: Other

The field of connectomics aims to map structural and functional connections in the brain
by recreating 3-D models (connectomes) of brain regions using 2-D imaging techniques obtained through electron microscopy. Segmentation algorithms may be implemented to realign images and map synaptic connections in these 3D models; however, due to the incredibly small scale of synaptic connections, these algorithms often face a multitude of unavoidable errors which must be identified and corrected through the efforts of artificial intelligence and human proofreaders.”Ground truth” is typically determined by expert proofreaders: proofreaders with an extensive education in neuroscienceā€”but the number of expert proofreaders available is certainly limited. Large cohorts of new proofreaders are greatly needed, but training can be difficult and timeintensive, leading to a necessity for both assessing proofreader performance in an accurate, streamlined manner as well as determining new ground truth in a reliable manner. Current methods for deciding ground truth and evaluating proofreaders can be biased and provide an inaccurate portrayal of proofreaders. Our Proofreader Evaluation Analysis toolKit (PEAK) proposes a continuously learning algorithm which provides unbiased reflections of individual performance relative to other proofreaders, as well as a reliable way to determine new ground truth, independent of expert proofreaders past the first iteration.

Keywords: connectomics, proofreading, Bayesian inference, performance evaluation, Kernel Density


  • Fadil Santosa


  • Justin Joyce
  • Victoria Rose
  • Will Gray Roncal

Team Members

Project Links

The MICrONS Minnie Dataset is a publically accessible functional connectome. It contains 200,000 cells, 75,000 neurons with physiology, and 523 million synapses. This 3d segmentation generations of the data are shown in the visualization.

< Back to All Projects