Engineering – Page 16 – UROP Spring Symposium 2021

Engineering

Head CT image analysis for detecting edema

Cerebral edema, which is swelling in the brain, is commonly found in patients suffering from head trauma, injuries, or other diseases and can be fatal. The purpose of this research project was to develop a method for automatically detecting and segmenting edema in head CT scans in order to make it faster and easier for clinicians to diagnose and treat traumatic brain injury (TBI) patients. However, edema is difficult to segment due to its unclear boundaries and its similarity in pixel value to other brain tissue. In previous research, most methods for segmenting edema have either been semi-automated or for MRI scans. More accurate methods require MRI scans, but even though an MRI scan is more detailed and can make it easier to segment edema, CT scan is the gold standard for evaluating brain injuries and is faster and more widely available. Therefore, automatic segmentation of CT scans will be very beneficial. In this project, the active contours without edges method, developed by Chan and Vese, is used with manually segmented hematoma as the initial contour. The method was developed in MATLAB. The segmented edema was then compared with the manually segmented images, and the DICE score was used to measure the accuracy. Currently, this method successfully segments select CT scans. However, it needs improvement in order to be more generalizable. In the future, I will look further into other techniques such as deep learning that can help improve accuracy and generalizability in automatic edema segmentation.

Developing Fast and Unbiased Computer Vision Algorithms

The research project I am participating in is “Developing Fast and Unbiased Computer Vision Algorithms” through the Multidisciplinary Design Program and the Transportation Research Institute. We’re trying to make a computer vision algorithm that could detect if drivers are paying attention to the road or distracted such as being on their phones. The algorithm itself should be as efficient and reliable as possible. To get our results, we look at frames of videos of people driving and create data sets and coding logs based on what the driver is doing. We have multiple people log the videos to create a benchmark of what the driver is doing. We also change our operational definitions of what we are looking for in the videos. The coding logs give us a benchmark for the algorithm so it can accurately judge what actions are distracted driving. By changing the variables we’re analyzing and improving the benchmark, we can make the algorithm more efficient, especially when we have a lot of different types of videos with varying lighting, subjects, and difficulty. Our research is vital because while transportation safety is important and an accurate algorithm detection distracted driving could help reduce the number of car accidents and car deaths, on a larger scale, our improvements of this computer vision algorithm would help improve how computer vision algorithms are created and applied in general. Overall, computer vision algorithms have shown to be biased especially with variables like skin color, sex/gender expression, and lighting. Those variables negatively affect the accuracy of the algorithm. Through our research, we could use our same methods and data to help other computer vision algorithms become more accurate and efficient.

Pancreas Segmentation via Image Processing and Machine Learning

Pancreas segmentation is consistently one of the least accurate out of all organ segmentation due to its low contrast in its boundary and high anatomical variability in its geometric properties. Many hospitals and biomedical laboratories therefore do not have accurate methodologies for cancer detection, 3D modeling, etcetera for pancreas scans. This paper serves to improve current pancreas segmentation methods by making use of a type of Convolutional Neural Network (CNN) called a U-Net in addition to image processing techniques such as an Integrated Hausdorff-Sine Loss Function for preprocessing and 3D Gaussian smoothing for post processing. The CNN was trained on Computed Tomography (CT) images in n-fold cross-validation from a public NIH dataset and a private dataset from Michigan Medicine. Data augmentation was performed on the CT scans using Keras, an open-source neural network library. Accuracy and precision of the CNN was tested using a Dice-Sørensen Similarity Coefficient (DSC) and Jaccard Index (JI). The mean DSC and JI achieved for the proposed methods are expected to be at least 70.00%.

Examining Kinematic Differences of Vestibular Gait

The vestibular system is a sensory system found in the inner ear that provides the brain with information about motion and spatial orientation. It helps maintain balance, stabilize the head/body during movement, and maintain good posture. If the vestibular system is impaired, a patient may experience dizziness, imbalance, and difficulty walking. Vestibular disorders are difficult to diagnose as they can stem from a variety of underlying conditions and require specialized equipment to ascertain a diagnosis (e.g., dynamic posturography, rotational chair). The long-term goal of this study is to use kinematic data to develop gait assessment technology that can automatically identify individuals with vestibular disorders. In this project, we used motion capture technology (MOCAP) to collect and analyze full-body kinematic data from subjects with vestibular issues and healthy controls to better understand kinematic differences between healthy and vestibular gait. Twenty-one subjects (11 with vestibular deficits and 10 healthy) were recruited to participate in the walking study. The subjects performed a series of tasks that required them to walk with varying speed and constraints (eyes closed, walking backwards, etc.) while wearing a set of full-body MOCAP markers. The data collected was then used to compute descriptive gait metrics, including step length, step time, and head/trunk rotations, using a motion-analysis software called Visual3D. Inferential statistic tests (t-tests) were then performed to compare the various metrics of healthy individuals with subjects with vestibular deficits. Conclusions have not yet been drawn as the study is still in the data processing phase. However, we plan to leverage gait metrics that show statistically significant differences between the two groups in order to develop automatic vestibular gait detection algorithms.

The relationship between body shape and air-breathing organ morphology in Anabantarian fishes as revealed by morphometrics

The Anabantarians are a diverse group of tropical, freshwater fishes which use air-breathing organs (ABO) within their skulls to live in harsh environments (e.g., warm, hypoxic pools). Space is limited within the skull however, suggesting that evolutionary changes to ABO size and shape can influence neighboring skeletal structures, a concept called morphological integration. We conducted this study to determine if there is a correlation between the modularity of fishes and the presence of an Air-Breathing Organ (ABO). The discovery of how the introduction of a new structure within a limited space, such as the skull, affects the relationship between neighboring structures, can help the science community better understand what affects integration between structures. While conducting this study we examined all Anabantarian families (e.g., Anabantidae, Osphronemidae, Badidae). We hypothesized that the presence of an ABO would create a more integrated cranial skeleton, and as a result decrease the modularity of the fish overall. To test this, fish skeletons were visualized through x-ray imaging, and aspects of fish phenotype were measured using linear morphometrics. We compared body shape metrics that correspond to aspect ratio, cranial elongation, and post-cranial elongation among fishes with and without ABOs. Our findings demonstrated that air-breathing organs have evolved 2-3 times, independently across anabantarians, and although fishes with ABOs are more diverse than fishes without, these findings were not statistically significant. Relatedly, there were no significant differences in body shape metrics among air-breathing and non- air-breathing fishes; however, the body shapes of fishes with ABOs evolve as much as 3.5 times faster than those fishes without.

lsa logoum logo