Colloquia Announcements

Wednesday, 4th March 2020, 5:05pm, MM217

Screening for prostate cancer reduces prostate cancer mortality; however, it leads to oversampling, overdiagnosis and overtreatment of indolent cancer. Both randomized trials and long-term observational data support active surveillance (AS) as a safe alternative to immediate treatment in appropriately selected patients. However, as emerging data show an increased risk of metastasis for patients with delayed treatment, there are increasing concerns as to the proper selection of patients for AS. Prostate cancer'???'s multifocal and highly heterogeneous nature is the Achilles heel of effective patient risk stratification, and standard ultrasound-guided biopsies have proven to be inadequate as a measure of sampling for clinically significant prostate cancer. In imaging, this heterogeneity is mapped using the different sequences in Multiparametric MRI (mpMRI) of the prostate, or so called "habitats" - areas with different risk for aggressive cancer. We have developed a quantitative mpMRI pixel-by-pixel habitat risk score (HRS) platform, which is superior at identifying high-grade disease of Gleason score (GS) 3+4 and higher based on a prostatectomy cohort in ROC-AUC analysis. HRS is displayed as a 3D heat-map and implemented in the biopsy workflow to assure that the most aggressive habitat of the tumor is targeted.

This is another in the Department of Computer Science Pizza Seminar Series. Refreshments will be served at 4:30pm outside MM217.

Wednesday, 26th February 2020, 5:05pm, MM217

Land Access for Neighborhood Development (LAND) is web-based mapping tool built to help visualize the distribution of local institutional, government-owned vacant and underused properties. Developed through a collaboration between the UM Office of Civic and Community Engagement and the UM CCS (IDSC) with support from Citi Community Development (Citibank), LAND enables policymakers and community-based organizations to identify potential affordable housing development opportunities in transit-served ares. A multi-tier application, LAND is built using a combination of Javascript (AngularJS), HTML, CSS (bootstrap), node.js., Apache Solr and PostGIS. The presentation will address both the need/usecase for LAND as well as its architecture.

This is another in the Department of Computer Science Pizza Seminar Series. Refreshments will be served at 4:30pm outside MM217.

Wednesday, 19th February 2020, 5:05pm, MM217

Current testing of ad effectiveness rely on methods such as content analysis, manual content annotation, focus groups, and opinion surveys. These methods are costly, time consuming and lack immediacy. We propose a computational approach that uses deep neural representations of visual, audio, and natural language content of ads to predict viewers' responses. Our approach employs the content as input and does not require the manual specification of content features as needed by conventional machine learning. We develop our approach on YouTube video ads and to build a predictive model of viewer response we use the YouTube engagement data (like/dislike, comment sentiments), as well as ratings by human coders. We find that of the three media data - video, audio and text - the video data is the most predictive of viewer response. This signifies that visuals contain the main predictive information for ad response. We also find that including computationally extracted multimedia timelines akin to the narrative timelines of stories improves the prediction of viewer response and that most multimedia timelines do not induce brand purchase intentions. The main implication of our work is that computerized systems can complement and perhaps even substitute current methods for providing viewer feedback on ad effectiveness.

This is another in the Department of Computer Science Pizza Seminar Series. Refreshments will be served at 4:30pm outside MM217.

Wednesday, 12th February 2020, 5:05pm, MM217

The National Weather Service, the National Hurricane Center, and other organizations regularly publish forecasts of where tropical storms and hurricanes may go and cause damage. However, a growing body of evidence suggests that most people can't read those visual products - maps, graphs, etc. This talk describes an effort to understand where misunderstandings of hurricane products may come from, and what we can do to address them.

This is another in the Department of Computer Science Pizza Seminar Series. Refreshments will be served at 4:30pm outside MM217.

Wednesday, 5th February 2020, 5:05pm, MM217

Artificial intelligence (AI) has great potential to augment the clinician as a virtual radiology assistant (vRA) through enriching information and providing clinical decision support. Deep learning is a type of AI that has shown promise in performance for Computer Aided Diagnosis (CAD) tasks. A current barrier to implementing deep learning for clinical CAD tasks in radiology is that it requires a training set to be representative and as large as possible in order to generalize appropriately and achieve high accuracy predictions. We present an Active Semi-supervised Expectation Maximization (ASEM) learning model for training a Convolutional Neural Network (CNN) for lung cancer screening using Computed Tomography (CT) imaging examinations. Our learning model is novel since it combines semi-supervised learning via the Expectation-Maximization (EM) algorithm with active learning via Bayesian experimental design for use with 3D CNNs for lung cancer screening. ASEM simultaneously infers image labels as a latent variable, while predicting which images, if additionally labeled, are likely to improve classification accuracy. The performance of this model has been evaluated using three publicly available chest CT datasets: Kaggle2017, NLST, and LIDC-IDRI. Our experiments showed that ASEM-CAD can identify suspicious lung nodules and detect lung cancer cases with an accuracy of 92% (Kaggle17), 93% (NLST), and 73% (LIDC) and Area Under Curve (AUC) of 0.94 (Kaggle), 0.88 (NLST), and 0.81 (LIDC). These performance numbers are comparable to fully supervised training but use only slightly more than 50% of the training data labels.

This is another in the Department of Computer Science Pizza Seminar Series. Refreshments will be served at 4:30pm outside MM217.

Monday, 3rd February 2020, 11:15am, UB230

Vision researchers have been analyzing behaviors of radiologists during screening to understand how and why they miss tumors or misdiagnose. In this regard, eye-trackers have been instrumental in understanding visual search processes of radiologists. However, most relevant studies in this aspect are not compatible with realistic radiology reading rooms. In this talk, I will share our unique experience for developing a paradigm shifting computer aided diagnosis (CAD) system, called collaborative CAD (C-CAD), that unifies CAD and eye-tracking systems in realistic radiology room settings. In other words, we are creating artificial intelligence (AI) tools that get benefits from human cognition and improve over complementary powers of AI and human intelligence. We first developed an eye-tracking interface providing radiologists with a real radiology reading room experience. Second, we proposed a novel computer algorithm that unifies eye-tracking data and a CAD system. The proposed C-CAD collaborates with radiologists via eye-tracking technology and helps them to improve their diagnostic decisions. The proposed C-CAD system has been tested in a lung and prostate cancer screening experiment with multiple radiologists. More recently, we also experimented brain tumor segmentation with the proposed technology leading to promising results. In the last part of my talk, I will describe how to develop AI algorithms which are trusted by clinicians, namely “explainable AI algorithms". By embedding explainability into black box nature of deep learning algorithms, it will be possible to deploy AI tools into clinical workflow, and leading into more intelligent and less artificial algorithms available in radiology rooms.

Wednesday, 29th January 2020, 5:05pm, MM217

Due to the negative environmental impact of damaging organic solvents and the high-cost of chemical waste disposal, the search for alternative, renewable solvents is a top priority in the chemical industry. Part of a worldwide emphasis on improving sustainability, the push has led to increasing interest in a set of environmentally-friendly solvents: ionic liquids (ILs). These solvents, primarily composed of ions, e.g., [RMIM][PF6], provide numerous advantages from both chemical and environmental perspectives. Despite the excitement surrounding these solvents, a fundamental understanding of how they operate is still lacking. This seminar will highlight our efforts developing and applying computer models to both elucidate how these solvents function and to design the next generation of cost-effective ???green??? solvents. Specific topics to be covered include force field parameterization, the development of machine learning neural network software for simulating chemical reactions, and the creation of machine learning genetic algorithms for automated force field development.

This is another in the Department of Computer Science Pizza Seminar Series. Refreshments will be served at 4:30pm outside MM217.

Wednesday, 22nd January 2020, 5:05pm, MM217

Modern oscilloscopes, data loggers, and related electronic measurement instruments generate a large amount of waveform data for many long-duration waveform capturing and analysis tasks. The contrast of time scales of long-duration waveform capturing (e.g., signal length of hours in high sampling rate) and analysis (e.g., signal fragments of several microseconds) produces unique big data challenges for signal acquisition, storage, organization, analysis, and user interactions. We have designed and implemented several long-duration waveform processing algorithms for waveform data organization, analysis, visualization, and user interaction at big-data scale. To cope with the real-time processing and user interaction demands within the hardware constrains of instrument platforms, the proposed algorithms utilize multiple layers of test signal synthesis, data pre-sorting, database query, and time series similarity quantification at various speed-precision tradeoffs. We also integrated typical instrument usage patterns into the design process to provide intuitive user interfaces for efficient data organization, flexible data record navigation, and in-depth experimentation.

This is another in the Department of Computer Science Pizza Seminar Series. Refreshments will be served at 4:30pm outside MM217.

Thursday, 9th January 2020, 11:15am, UB230

Modern big data analytics are typically constructed through a multi-step procedure, often referred to as a "workflow", with each step in a workflow representing a specific algorithm or function. Over the years, many scientists have been frequently recreating their own data analytics algorithms from scratch. This results in a major duplication of effort, and it wastes money, time and other important resources. In this talk, I will discuss the necessity of building a data science infrastructure capable of recommending reusable algorithm resources (artifacts) when scientists design big data analytics workflows. I will illustrate an application of machine learning and information retrievalmethods to tackle the automatic annotation problem. Aiming to highlight the uniqueness of an artifact, automatic annotation is an important problem in data science, since descriptions from contributors might be subjective and not kept up to date. Our work illustrates the use of a generative model that learns from artifact usage provenance to refine artifact representations, and uses these learnings to support a category-aware artifact recommendation method.

Previous Colloquia Announcements