These projects were completed during graduate school.
Metal Artifact Reduction in Computed Tomography
Jan ‘17 - March ‘17
This project was completed for a graduate biomedical signal processing course. We investigated metal artifact reduction (MAR) methods used in x-ray computed tomography (CT). Motivation: The filtered back projection (FBP) algorithms used to reconstruct CT images are based on the monochromaticity assumption of the x-ray source, such that the sinogram (Radon transform data) is an approximation of a material’s attenuation coefficient distribution at a fixed energy level. For most soft tissues in the body, the monochromaticity assumption is valid; however, if metal is present in the structure, the assumption is violated because the attenuation coefficients of metallic materials vary greatly with energy level, thereby leading to a significant mismatch between the sinogram data and the Radon transform of the image. This produces ‘star artifacts’ in the reconstructed image that result in a loss of anatomical information. Metallic objects in patients often include dental implants, surgical clips, steel-hip prostheses, bullets, or shrapnel.
We developed mathematical models and simulations for CT using an (i) absorption-only (conventional) model and (ii) full absorption and scattering model. We replicated existing algorithms for MAR, namely the sinogram interpolation method and a gradient-descent-with-image-segmentation algorithm. Lastly, the benefits of non-local means prefiltering in conjunction with sinogram interpolation were explored. Project member names removed for privacy
Robotic Object Tracking for Path Planning Using K-means and OpenCV
Sept ‘16 - Dec ‘16
This project was completed for a graduate machine learning course. Using a Raspberry Pi, a small mounted camera, an Arduino microcontroller, and an iRobot Braava, my team built an autonomous robotic system that executed a set of maneuvers based on visual input and classification algorithms programmed into the Raspberry Pi. Real self-driving car systems are, obviously, much more complex (i.e LIDAR, GPS, etc.) and span varying levels of autonomy (see SAE’s six level classification system), but our objective was to build a simplified model to explore the efficacy of various algorithms. We focused on k-means, Canny edge detection, and algorithms from OpenCV (Open Source Computer Vision). Canny edge detection and OpenCV were used in image segmentation, while k-means was used for object classification.
Automated QA of human brain MRI image segmentations using SVM
Jan ‘16 - Aug ‘16
This was a thesis level project completed for the Gordon Institute of Engineering Leadership. The objective was to train a binary classifier to assess the quality of brain MRI segmentations. The current methodology for QA of MRI segmentations, worldwide(!), is manual visual checking which is incredibly tedious, time consuming, and error-prone. To this date and to my knowledge, very few work, if any, has been done in this area. For this project, I focused primarily on the SVM classifier, spending a large amount of time on exploratory analyses such as feature selection and feature engineering. The report goes into much more depth (warning: it’s long). Some names are removed for privacy. (Looking back, I should have written it in LaTeX…)