Medical image segmentation modeling is a high-stakes task where direct communication and interpretation of uncertainty is crucial for addressing visual ambiguity. Prior work has developed segmentation models utilizing probabilistic or generative mechanisms to infer uncertainty from labels where annotators draw a singular boundary. However, as these annotations cannot directly represent an individual annotator's uncertainty, even specialized models trained on these standard representations produce uncertainty maps that are difficult to interpret. We propose a novel segmentation representation, Confidence Contours, which uses high- and low-confidence ``contours’’ to capture uncertainty directly, and develop a novel annotation system for collecting contours. We collect both standard and Confidence Contours annotations on the Lung Image Dataset Consortium (LIDC) and a synthetic dataset simulating the structural ambiguity of many medical segmentation problems, FoggyBlob. Our analysis show that Confidence Contours provide high representative capacity without requiring significantly higher annotator effort. Moreover, general segmentation models trained on Confidence Contours can produce significantly more interpretable uncertainty maps than models with specialized mechanisms for uncertainty, and they can learn Confidence Contours at the same performance level as singular annotations. We conclude with a discussion on how we can infer regions of high and low confidence from existing segmentation datasets. Our data-centric approach crucially brings attention to the importance of human factors in responsible and robust AI, which have often been overlooked in model-centric medical segmentation work. By troubling and rethinking the very way that the ground truth is represented, our work opens up new paths of inquiry towards more human-friendly models -- paths which begin from the data.

Current smart home devices understandably prioritize the needs of a primary user/owner, who is also typically the purchaser of the device and a corresponding subscription plan. Yet smart home cameras and other smart devices with microphones, location tracking, and other spatial sensing capabilities invariably impact the privacy of people nearby, such as family, friends, guests, neighbors, and domestic workers. We refer to these affected nearby people as adjacent users (or adjacent subjects) because they may interact with smart devices but do so with relatively little or no direct awareness, consent, access, control, or benefit. Although work in privacy and security has begun to address the privacy needs of adjacent users, there is little design research that has responded with either concrete interventions like proposals or prototypes. We present a novel speculative design of a smart home camera called Arca with a physical camera prototype and a mobile application. A significant insight of our empirical and design research is that the most common issues with smart camera privacy is the interpersonal tensions and conflicts stemming from inadequate disclosure, consent, autonomy, and transparency from primary owners. Whereas traditional privacy/security research often focuses on harms from improper disclosure of personally sensitive information, our research suggests that many adjacent users do not necessarily mind being recorded, they do mind the lack of “communication,” “respect,” and “professionalism” from primary users. Furthermore, our studies reveal that even if our specific privacy modes and access sharing features are not regularly used, they may nonetheless function as mechanisms to facilitate better, more open conversation between primary and adjacent users. We continue our work with the goal to enhance adjacent user privacy and experience with privacy-sensitive camera features and reduce tension between adjacent users and primary users.

Ophthalmic drug administration has been increasingly prevalent in recent years, with eyedroppers being utilized to administer costly medication like that for glaucoma. There haven’t been many solutions addressing eyedropper instillation for those with preexisting conditions like arthritis, who often deal with a host of problems when administering them: producing the necessary force to distill a drop, aiming the drop into the eye, and contamination of the eyedropper tip. We are testing the question of whether accessible eye drop aids can significantly improve eyedrop compliance and distillation for the elderly. Solutions to eye drop administration can save money and make the overall process easier for many patients. Existing solutions on the market seem to address the issue of contamination using apparatuses that press onto the lower eyelid, but there is still much to be desired with the force and aim required. Many require the use of gripping or squeezing, motions that many elderly patients can’t apply as much force with. I propose a couple of solutions to these problems in the form of eyedropper aids that each make use of a few different methods, including translating the motion, applying the force with different limbs, and even mechanizing the force required. Through a quantitative study, I hope to eventually test these prototypes through an ophthalmology clinic among a wide variety of elderly. Assessing these prototypes through both questionnaires and observation, I hope to notice an increase in effectiveness from previously existing apparatuses. We will use a survey to ask a variety of questions to around 100 elderly patients with varying expertise in eye drop instillation. The survey will ask whether the tool was more useful, easier, how hard it was to assemble, and we will also monitor quantitatively whether the accuracy of drops actually instilled was better. This work hopefully saves patients money from medication cost from a reduction in wastage, allows for better administration of medicine, and eases the process of distillation.

Foams are used on a daily basis in many different parts of ordinary life ranging from mattresses to automobiles. Yet, the technology for creating variable-density foams has not progressed beyond the method of gluing together sheets of foams with different densities. We present a new method of making variable-density foams using a 3D process that takes advantage of Viscous Thread Instability, a phenomenon that causes viscous liquids to coil when extruded out of a small opening. An everyday example of this would be the way honey coils as it is being poured onto a delicious treat. Similarly, when filament is extruded from a 3D printer, it melts into a semi-viscous liquid and coils in a stochastic process. Viscous Thread Printing (VTP) can be used to manufacture foam-like objects using standard Fused Deposition Modeling (FDM) 3D printers. We demonstrate a method to predictably print these foams by using a custom slicing software that produces VTP-specific instructions for 3D printers. The slicer can generate arbitrarily shaped VTP foams and manipulate the coiling characteristics of the filament, which allows for the production of foams with varying stiffness within a single print - something that traditional foam manufacturing methods cannot handle. As such, Viscous Thread Printing could be a viable alternative to traditional foam manufacturing, as our new slicer reliably generates variable-density foams in a more efficient way.

Hearing loss is a prevalent problem throughout the world, and the process of getting and fitting hearing aids is becoming more accessible to meet this demand with the emergence of technologies that allows a hearing aid user to self-fit their devices using a smartphone application to meet their individual needs. Our lab has developed two different self-fitting applications that vary in their designs of the graphical user interfaces. The goal of the current study is to conduct systematic evaluations of the applications with older-adult participants. A group of older adults were recruited. Following hearing assessment and cognitive screening, these participants were instructed to complete self-fitting using the two applications, in random order. For each application, the self-fitting procedure involved adjusting the hearing-aid settings interactively while listening to a continuous speech presented together with a background noise through the hearing aid. The procedure was repeated for several background noises, representing different acoustic scenes (outdoor, restaurant, etc.). Following hearing-aid fitting, the participants were surveyed about their perception on the usability of the applications and preferences between the competing user interface designs. Additionally, all participants rated on speech quality through hearing aids with three different settings, including the self-selected settings using the two applications. The quality rating data will be analyzed with the participant’s age, cognitive status, and comfort level with mobile technologies controlled for. We hypothesize that older adults may have preferences for the user interface that is more functional and visually simplistic, and that the user’s preferred hearing-aid settings would change with the acoustic scene. The feedback we receive from this project is a crucial component of the whole application design process, and if our hypotheses are supported, it will reveal potential efficacy and direct future areas of improvement for our final applications.

The algorithmic design of traveling salesman problem (TSP) is one of the most famous graph based problems. With recent developments, one approximation algorithm for the asymmetric case of this classical problem became the milestone in the field due to its novel application of modern continuous optimization techniques onto discrete mathematical objects. The purpose of our project is to find the probability distribution that maximizes randomness (max entropy distribution) over a rooted and directed version of spanning trees arborescences). Pprevious work shows that max entropy distribution over undirected spanning trees is essentially the uniform distribution, which makes spanning tree sampling extremely fast. Our goal is to find whether the max entropy distribution over arborescences assume similar convergence behaviors. We hypothesize that through convex programming formulation, the eventual outcome of the max entropy distribution of arborescences is also the uniform distribution, due to structural similarities between the two objects. However, the final result we arrived at asserts that the uniform distribution over arborescences from a graph does not maximize randomness. Based on this finding, we further compared other behaviors of arborescences against spanning trees, and through the discovery of graphic examples, we found out that arborescences essentially fail to possess the concentration properties known for spanning trees. Therefore, our work aims to further motivate for a generalized explanation behind such distinct behaviors of mathematical objects. Through extending the probabilistic lens onto directed versions of well-studied graph structures, we hope the new techniques based on their properties would lead to future algorithms that factor in the real world complexities associated with cost or distance asymmetry, one example would be the asymmetric costs of traveling between two cities. In the long run, we hope to develop more robust algorithms with less reliance on ideal mathematical conditions in market operations and data analysis research.

Smartphone detection of anemia using patient photos has the potential to provide a non-invasive method of measuring hemoglobin levels, introducing the possibility of increasing the accessibility and cost-effectiveness of current practices. While traditional methods of anemia detection require a complete blood count by a trained healthcare professional, smartphone detection instead relies on the user to take a high quality picture of their fingernails. However, it currently lacks the ability to provide feedback to the user on the quality of their image. For example, an overexposed image or one with low fingernail visibility can lead to inaccurate predictions of hemoglobin levels. We propose that machine learning classification methods can analyze these patient images to estimate the image quality and predict the effectiveness of smartphone detection of anemia for a given image. With various classical machine learning models, we demonstrate and compare the capabilities of each in classifying images of patients’ hands as being of “good” or “bad” quality (or on a more granular numerical scale) when given features of the images. Preliminary results show that a logistic regression model reaches 91.4% accuracy labeling images when compared to empirically assigned labels, and we expect iterative models to achieve improved performance. When completed, we would propose that this classifier could be used in the field to identify if patient image is of high enough quality to produce an accurate measurement of hemoglobin levels in real-time, providing feedback on the phone to adjust or correct the image-taking process.