Sophia Scarano

Hybrid Quantum-Classical Biomedical Image Segmentation Network (ongoing)

Medical image segmentation has an essential role in computer-aided diagnosis systems for many different applications. Here, we look at the application of image segmentation on brain MRI images for brain tumor detection, using the KiU-Net architectural framework. In this case, KIU-Net is preferable over the industry standard U-Net because of its over-complete representation methods for more accurate biomedical image segmentation of small and detailed objects such as brain tumors. Additionally, the implementation of Hybrid Quantum-Classical Convolutional Neural Networks is embedded into the feature map creation step in each CNN layer within the KiU-Net framework. Because quantum computing has demonstrated to be able to output complex wave functions with a few number of gate operations, the adaptation of quantum simulations for matrix operations has been adapted to enhance the feature mapping process, which is the most computationally intensive part of CNNs.
To learn more about this project, click here.

Energy Deposition Image Classification (ongoing)

The goal of this project is to perform multi class object detection, comparing the difference between using a standard convolutional neural net and using the more novel capsule network approach for this task. Additionally, I would like this project to eventually be one where I can test different model interpretability algorithms such as LIME, SHAP, and DASP, all state-of-the-art methods. I am also interested in seeing how a Saliency Map might be beneficial in visualizing these model interpretations.
This project is ongoing, and right now I am working on training a custom YOLOv4 object detector using Alexey Bochkovskiy's Darknet framework. I will be using the YOLOv4 model as my convolutional neural net framework because of it’s generalization capabilities which are useful for custom multi-class detection, and its real-time processing benefits which are useful in a scenario where the incoming TPC data must be processed and classified quickly.
Additionally, I plan to test both the Matwo-CapsNet method and the CapsNet-Keras method for this dataset. Of course, I wish to see if either outperforms the YOLOv4 convolutional neural net. However what I am most interested in seeing is if the ‘capsule’ aspect of the network can allow for a further interpretation and classification of the decays happening within the chamber itself.

Handwritten Equations to LaTeX

The end result of this project is an app that takes an image of a handwritten equation, and after prompting the user to follow certain instructions, generates the associated LaTeX syntax for the handwritten equation. Used computer vision (CV) and compared “You only look once” (YOLO) and GoogLeNet real-time, object detection and classification performance for image classification. All code was executed using Google CoLab. For training, I identified 10,000 calculus equations and trained CNN enabling custom object detection and multilabel classifications for 74 different classes. Each class is a mathematical expression ('x', 'lim', '2', etc.) that are the most commonly found in handwritten Calculus limit equations. Unfortunately, the 74-class custom object detector crashed after running for over 100 hours in CoLab, so I am currently thinking of ways to resolve this issue. However, this project is not my main goal at the moment so it may be a while until a fully functional app is released.
To learn more about this project, click here.

California COVID-19 and Wildfire Amenity Avoidance

The end result of this project is an app that helps users avoid COVID-19 hotspots and dangerous ongoing wildfire locations in California. This app allows users to input their current location, desired location, and amenity they are interested in visiting en route. Based on an additional input of the distance they are willing to deviate from their route, this application displays all location of their desired amenity in that area, along with each location's assiciated "risk level" for both COVID-19 and wildfire proximity. To learn more about this project, click here.

Booz Allen Hamilton

Only non-PhD team member in this role. Variety of technical responsibilities for both internal and client-facing positions. Explanation limited due to Booz Allen Ethics and Compliance release of information guidelines.

Experienced in leading small software teams, working collaboratively with peers in large project teams, and in conducting independent, results-driven quantum application projects.

General Assembly

480-hour masters-equivalent immersive program applying data collection and cleaning, analysis, modeling, data visualization, and machine and deep learning techniques to solve real-world data problems.

Completed challenging application process and earned admission to exclusive, rigorous, world-class data science immersive program. Applied linear algebra, statistics, and complex mathematics concepts to client and commercially viable applications addressing current issues using real-world data. Used image and natural language processing, deep learning techniques, and multiple programming languages in developing functional solutions. Projects completed autonomously and in collaborative, cloud-based environment.

Joint Quantum Institute

Led small team and completed autonomous hardware and software projects contributing to trapped ion-photon quantum information system.

Maryland Nanocenter

Spent 6 months contributing to ongoing Ultrafast Optical Science project.

University of Maryland