I was supported by NSF IPAN to work with Dr. Linqing Feng at KIST. We were interested in automating cell segmentation to study micro-scale brain connectivity. This is crucial to efficiently analyze large volumes of electron microscopy (EM) images of regions such as the hippocampus. To test the efficacy of such a method, I implemented TensorFlow code for the DCAN model on human U2OS cell images.
DCAN by Chen et al. builds upon the fully convolutional network (FCN), in which semantic insight of deeper layers is combined with locality details of shallower layers. This end-to-end network can make per-pixel predictions for tasks like semantic segmentation.
Time-evolving graphs can be observed in a variety of domains, such as neuroscience and sociology. In connectomics, we can infer temporal networks from fMRI data; the nodes are voxels (volume units of neurons) and the edges are their thresholded associations. How can we track the evolution of domain-specific communities in such graphs?
ECOviz is a system that summarizes and visualizes temporal network changes in a semi-supervised manner. Due to the small-worldness of brain networks, we used a subset of labeled nodes to inform TimeCrunch, a dynamic graph summarization algorithm. In addition to showing the topology of summary structures, ECOviz allows users to compare the effects of preprocessing parameters (i.e., threshold and time interval granularity).
Advisor: Andrew DeOrio
Collaborators: Ajaay Chandrasekaran, Edmond Cunningham, Levin Kim, Wenlu Yan, Xinghai Zhang, Yaman AbdulhakPoster
Sponsored by the Kellogg Eye Center, this was a two-semester project in the Multidisciplinary Design Program (MDP). Retinal dystrophies are inheritable disorders that require expensive genetic tests and medical expertise to diagnose. Given patient inheritance and history data, we built a web app with a data-driven model to predict retinal dystrophy diagnosis.
I contributed to the web app prototype to (1) collect patient data from a user and (2) visualize output from a predictive model built by my teammates. This model first predicted inheritance pattern, then paired this with patient history to output the final diagnosis via an RBF kernel SVM.
Under Prof. Jun Zhang and visiting faculty Yinbin Lei, this was an independent study project in the psychology department. Though algorithms to extract concept hierarchies exist, I contributed code to visualize and query the computed lattice through set operations. This work was motivated by analysis of concept formation data.
Formal concept analysis (FCA) is a method for deriving a concept hierarchy from a table of object-attribute relations. A formal concept is composed of a set of objects and their attributes. Concepts can be partially ordered into a lattice, or hierarchy, such that edges represent set closures.