node_title | node_title
Research Plan
node_body | node_body
- Implement proofreading tools to enable long-range segmentation-driven tracing. Currently, we cannot trace large volumes due to the size of the dataset. We need to modify our clients to enable more scalable tracing.
- Devise algorithms to fully automate synapse prediction. We believe that we can tolerate some errors in synapse annotation. Therefore, if we can improve the accuracy of prediction to 90 or 95 percent, manual annotation may be unnecessary.
- Use high-level priors such as neuron shape and graph connectivity to guide segmentation. Segmentation algorithms traditionally focus on boundary prediction to produce an over-segmentation. This over-segmentation is then improved using some sort of agglomeration routine. This process is flawed as it is possible for small segmentation errors to propagate and lead to catastrophic degradation in segmentation quality. We hope to focus more on high-level priors or hints that will better guide segmentation. By doing so, we might be able to make segmentation better algorithms that error more gracefully.
- Continue develop of reconstruction domain-specific API to facilitate better collaboration throughout the EM community.