BIDS Image Analysis Across Domains
BIDS ImageXD 2021
Dates: May 17-18, 2021
Timing: 10:00 AM – 2:00 PM Pacific
Watch this event: Event Videos
___________________________
Program Summary (Session titles link to full abstracts and speaker information below.)
Day 1: MONDAY, May 17
10:00 – 11:00 AM Pacific — Session 1: Global Change
- Welcome to BIDS ImageXD 2021 • David Mongeau, BIDS
- Session Introduction • Maggi Kelly, UC Berkeley
- Novel indicators of biodiversity contributions on ecosystem stability from multi-temporal remote sensing data • Iryna Dronova, UC Berkeley
- Hydrologic Applications of Data Assimilation • Manuela Girotto, UC Berkeley
11:00 AM – 12:00 PM Pacific — Session 2: Environmental and Social Justice
- Session Introduction • David Mongeau, BIDS
- ** Cancelled Presentation: Lakóta GeoSpatial Applications of the Black Hills Area - Makówapi Wítaya (Geography Together) • James J. Sanovia, Oglala Lakota College
- Image Integrity and Provenance in Critical Contexts: Human Rights, Deepfakes and Authenticity Infrastructure • Sam Gregory, WITNESS
- Online prescription fulfillment using Intelligence Document Automation • Anwitha Paruchuri, Accenture
12:00 – 1:00 PM Pacific — Session 3: Tools & Methods I
- Session Introduction • Alvin Cheung, UC Berkeley
- PyTorch3D • Jeremy Reizenstein, Facebook AI Research, London
- Drone Image Processing: Opportunities and Challenges • Avideh Zakhor, UC Berkeley
- Superpixels, Supervoxels, and Adaptels • Sabine Süsstrunk, EPFL
1:00 – 2:00 PM Pacific — ImageXD Community
___________________________
Day 2: TUESDAY, May 18
10:00 – 11:00 AM Pacific — Session 4: Materials Science
- Session Introduction • Daniela Ushizima, Berkeley Lab, UC Berkeley, UCSF
- GPU-accelerated Data Science with RAPIDS • Zahra Ronaghi, NVIDIA
- Space-Time Correspondence as a Contrastive Random Walk • Allan Jabri, UC Berkeley
- Quantifying solidification of metallic alloys with scikit-image • Marianne Corvellec, scikit-image core team and IGDORE; and C. Gus Becker, CANFSA
11:00 AM – 12:00 PM Pacific — Session 5: Health
- Session Introduction • Maryam Vareth, BIDS, UCSF, UC Berkeley
- High throughput lesion evaluation and quality control for incorporating quantitative imaging metrics into clinical practice • Anisha Keshavan, Octave Bioscience
- Pathology Image Analysis with Deep Learning • Saeed Hassanpour, Dartmouth
- Computational Image Phenotyping for Stroke and Prostate Cancer • Corey Arnold, UCLA
12:00 – 1:00 PM Pacific — Session 6: Tools & Methods II
- Session Introduction • Laura Waller, UC Berkeley
- The ImageJ Ecosystem • Kevin Eliceiri, University of Wisconsin-Madison
- napari: a fast, interactive, multi-dimensional image viewer for python • Nick Sofroniew, Chan Zuckerberg Initiative
- Automated quantification and analysis of human brain connections • Ariel Rokem, University of Washington
1:00 – 2:00 PM Pacific — ImageXD Community
Full Program (Pacific Time)
Day 1: MONDAY, May 17
10:00 – 11:00 AM — Session 1: Global Change
◦ Welcome to BIDS ImageXD 2021
David Mongeau
Executive Director, Berkeley Institute for Data Science (BIDS), University of California, Berkeley
◦ Session Introduction
Maggi Kelly
BIDS Faculty Affiliate; Professor of Environmental Science, Policy,and Management; and Faculty Director, Geospatial Innovation Facility; University of California, Berkeley
◦ Novel indicators of biodiversity contributions on ecosystem stability from multi-temporal remote sensing data
Iryna Dronova
Associate Professor, Department of Environmental Science, Policy & Management, Rausser College of Natural Resources, and Department of Landscape Architecture & Environmental Planning, College of Environmental Design, University of California, Berkeley
Abstract: Growing archives of remote sensing data provide a comprehensive outlook on dynamics of ecosystems in space and time and enable novel indicators of their stability and response to change drivers. Such indicators are especially appealing for detecting early signals of tipping points and biodiversity loss; yet, they are still under-developed for complex systems where signals of ecological processes are confounded by environmental variation and spectral background effects. This study explores the potential of remotely sensed indicators of ecosystem seasonality and its inter-annual stability to elucidate the contributions of plant diversity to long-term stability of ecosystem function using the example of wetlands as highly dynamic and spatially heterogeneous environments. Here we develop a new approach to characterize inter-annual consistency in ecosystem seasonality and demonstrate its sensitivity to vegetation diversity using a national-scale sample of 1,138 wetlands across the conterminous USA. We also identify a suite of important covariates that need to be considered in biodiversity-stability relationships at broad landscape scales. This analysis provides a new basis for assessing biodiversity-stability relationships at large geographic extents relevant to policy, conservation and planning and for developing new cost-effective indicators of ecosystem restoration and management outcomes.
◦ Hydrologic Applications of Data Assimilation
Manuela Girotto
Assistant Professor, Department of Environmental Science, Policy & Management, Rausser College of Natural Resources, University of California, Berkeley
Abstract: Accurate knowledge of spatial and temporal hydrological storages such as snow, soil moisture and terrestrial water storage are essential for addressing a wide range of important, socially relevant science, application and management issues. While in situ observational networks are improving, the only practical way to observe the land surface on continental to global scales is via satellite remote sensing. Though remote sensing can make spatially comprehensive measurements of various components of the land surface system, it cannot provide information on the entire system (e.g. deep moisture stores), and the measurements represent only a snapshot in time. Land surface process models may be used to continuously predict the temporal and spatial land system variations, but these predictions are often poor, due to model initialization, parameter and forcing errors. An attractive prospect is to combine the strengths of models and observations (and minimize the weaknesses of both) to provide a superior land surface state estimate. This is the goal of data assimilation. Data assimilation provides a better estimate of the environmental states than either models or observations could individually do. This presentation will focus on benefits and challenges of recent land surface data assimilation research efforts targeted at improving snow, soil moisture, groundwater, and terrestrial water storage hydrological states.
— Break (~20 mins)
Back to Program Summary
Day 1: MONDAY, May 17
11:00 AM – 12:00 PM — Session 2: Environmental and Social Justice
◦ Session Introduction
David Mongeau
Executive Director, Berkeley Institute for Data Science (BIDS), University of California, Berkeley
** Cancelled Presentation:
◦ Lakóta GeoSpatial Applications of the Black Hills Area - Makówapi Wítaya (Geography Together)
James J. Sanovia, Faculty - GIS Lab Manager, STEM Department, Oglala Lakota College
Abstract: The Black Hills (ȞeSápa) visualization mapping project is a collective of faculty-student research, various scientific communities, stories from our Lakóta-Dakóta ancestors, elders, and from the very few literature sources that exist that have any relevancy towards intergenerational knowledge. The imagery produced was aimed at bringing the Lakóta culture closer to the youth, the Lakóta Oyáte (people), and to Indigenous communities throughout the region. This work has been for educating the Oyáte and those willing to help so they too can learn why we continue to counter map, fight for decolonized mapping practices, the Black Hills, treaty territories, and for Grandmother Earth (Uŋčí Maká).
The Black Hills Red Race Track (ȞeSápa Kí Íŋyaŋka Očáŋku Šá) project, the most significant of these geospatial applications, is now in its 18th-year venture and still going. This was the first indigenous/cultural geography project that used modern geospatial technologies to tell the story of a Native American Tribe done by a tribal member. The overall premises of this project was a collective of several mapping endeavors over the years that range from using GIS, remote sensing, geology, geography, and Lakota culture. The ultimate significance of the Black Hills is that the Lakóta-Dakóta believe it is “The Heart of Everything that is”. Such an endeavor is only done by working together: Makówapi Wítaya (Geography Together).
Speaker Bio: Makówapi Wítaya Enterprises is a Lakóta based cultural mapping collective run by James Sanovia, Rosebud Sioux Tribal enrolled member (Sičáŋǧu Lakóta). Mr. Sanovia was born during the American Indian Movement where his mother comes from the Rosebud Sioux Tribe and his father from the Oglála Sioux Tribe, both in South Dakota. Mr. Sanovia gets his continuous support from his wife Lilly and their children. Mr. Sanovia has an AA in Pre-engineering from Oglála Lakóta College and a BS in Geological Engineering, and an MS in Geology and Geological Engineering from the South Dakota School of Mines and Technology. All three degrees emphasized cultural geospatial applications. James is nearing 20 years of cultural-science place-based mapping experience and 10 years as a faculty member at Oglála Lakóta College teaching geospatial and engineering-related topics. Through the guidance of both cultural and science advisors, James has been addressing environmental, cultural, and sovereignty issues through Indigenous-centered land mapping. Tašúŋka Witkó (Crazy Horse) is often quoted as saying, “My lands are where my dead lie buried. One does not sell the earth upon which the people walk”. James strives to center Indigenous lands through cultural mapping (counter-mapping) in support of decolonization initiatives. Through Makówapi Wítaya initiatives James also strives to give voice to all the two-legged, four-legged, winged ones, and all of plant life, thus, protecting Kéya Wíta (Turtle Island) with Indigenous values and virtues for ways of knowing to help move of forward in today’s world. Mr. Sanovia has been presenting Indigenous-centered mapping across the nation at tribal colleges, major universities, government entities, and to the public for almost two decades. It is time for Indigenous people to be the producers of maps of Indigenous places and the stories and history that goes with them and James has been leading such an initiative.
◦ Image Integrity and Provenance in Critical Contexts: Human Rights, Deepfakes and Authenticity Infrastructure
Sam Gregory
Program Director, WITNESS (WITNESS Media Lab)
◦ Online prescription fulfillment using Intelligence Document Automation
Anwitha Paruchuri
Artificial Intelligence Associate Principal at Accenture
Abstract: Unstructured medical prescriptions are continuously being processed as a part of business operations in the health care sector, and dedicated Subject Matter Experts (SME’s) spend considerable time and effort to read and interpret sequences of handwritten text with custom/domain specific vocabulary and manually extract clinical information from them such as Drug Name, Directions, PHI etc. As businesses are moving towards digital technology and AI based automation, there is a strong need for healthcare companies to leverage intelligent document understanding to increase the productivity for manual reviewers, scale to large customer volumes and improve the accuracy for prescription entering. The implemented solution adopts document classification, enhanced handwriting recognition, multi-modal embeddings and self-supervised learning, to scale and generalize across different layouts with limited amount of annotated data.
— Break (~20 mins)
Back to Program Summary
Day 1: MONDAY, May 17
12:00 – 1:00 PM — Session 3: Tools & Methods I
◦ Session Introduction
Alvin Cheung
BIDS Faculty Affiliate, and Assistant Professor of Electrical Engineering and Computer Science, University of California, Berkeley
◦ PyTorch3D
Jeremy Reizenstein
PyTorch3D Team, Facebook AI Research
I will introduce PyTorch3D as an efficient and accessible library for a range of 3D deep learning tasks including differentiable rendering. Then I will go on to describe neural rendering methods and our tools for experimenting with them.
◦ Drone Image Processing: Opportunities and Challenges
Avideh Zakhor
Chair & Professor, Electrical Engineering and Computer Science, University of California, Berkeley
Abstract: In this talk, I will give an overview of a reality capture pipeline for 3D building reconstruction from RGB imagery captured via an unmanned aerial vehicles. This is an important problem with applications in urban planning, emergency response, disaster planning, and building energy efficiency. We leverage the commercial software Pix4D to construct a 3D point cloud from RGB drone imagery, which is then used in conjunction with image processing and geometric methods to extract a building footprint. The footprint is then extruded vertically based on the heights of the segmented rooftops. Next I describe a method to estimate window to wall ratio (WWR), by applying semantic segmentation to the images containing windows and back-projecting them onto the 3D reconstructed building. Window to wall ratio has critical influence on heat loss, solar gain and daylighting levels with implications for visual and thermal comfort as well as building energy performance. Finally, I will describe a pipeline for semantic segmentation of 3D point clouds obtained via photogrammetry from aerial RGB camera images. Our basic approach is to directly apply deep learning segmentation methods to the very RGB images used to create the point cloud itself, followed by back-projecting the pixel class in segmented images onto the 3D points. This is a particularly attractive solution, since deep learning methods for image segmentation are more mature and advanced as compared to 3D point cloud segmentation. Furthermore, GPU engines for 2D image convolutions are likely to result in higher processing speeds than could be achieved using 3D point cloud data. We demonstrate results for two RGB Drone image datasets captured in Alameda, California.
◦ Superpixels, Supervoxels, and Adaptels
Sabine Süsstrunk
Professor and Head of the Image and Visual Representation Lab, Ecole Polytechnique Fédérale de Lausanne (EPFL)
Abstract: Superpixels provide a convenient primitive from which to compute local image features. They capture redundancy in the image and greatly reduce the complexity of subsequent image processing and computer vision steps. They have proved to be very useful for many tasks across multiple application domains, and are still very much in demand when deep learning techniques are not applicable. Size uniformity is one of the prominent features of superpixels. However, size uniformity rarely conforms to the varying content of an image. The chosen size of the superpixels therefore represents a compromise - how to obtain the fewest superpixels without losing too much important detail. We present an image segmentation technique that generates compact clusters of pixels grown sequentially, which automatically adapt to the local texture and scale of an image. Our algorithm, Adaptels, liberates the user from the need to choose of the right superpixel size or number. The algorithm is simple and requires just one input parameter. In addition, Adaptels are computationally very efficient, approaching real-time performance, and are easily extensible to three-dimensional image stacks and video volumes. We demonstrate that our Adaptels are superior to the respective state-of-the-art algorithms on quantitative benchmarks.
— Break (~12 mins)
Back to Program Summary
Day 1: MONDAY, May 17
1:00 – 2:00 PM — ImageXD Community
For this social and networking hour, participants will be able to engage directly in conversations and discussions with speakers and other participants. Explore the lounges to find or to initiate a discussion that resonates with your interests and research.
Day 2: TUESDAY, May 18
10:00 – 11:00 AM — Session 4: Materials Science
◦ Session Introduction
Daniela Ushizima
BIDS Research Affiliate, UC Berkeley; Staff Scientist, Computational Research Division, Berkeley Lab; Affiliate Faculty, Bakar Computational Health Sciences Institute (BCHSI), UCSF
◦ GPU-accelerated Data Science with RAPIDS
Zahra Ronaghi
System Software Manager, NVIDIA
Abstract: RAPIDS is a collection of open-source libraries for accelerating data science pipelines on GPUs, and is designed with familiar APIs for data scientists working in Python. This talk will cover an overview of this platform, core libraries including cuDF (GPU-accelerated dataframes) and cuML (GPU-accelerated machine learning), and Dask on GPUs for large-scale data analytics.
Speaker Bio: Zahra Ronaghi is a system software manager on the AI Infrastructure team at NVIDIA. She is primarily focused on GPU-accelerated machine learning and integration of RAPIDS libraries with cloud service providers and ML platforms. Prior to joining NVIDIA, Zahra was a postdoctoral fellow at Lawrence Berkeley National Laboratory (NERSC), where she worked on performance optimization of a tomographic reconstruction code and deep neural networks for neutrino telescopes.
◦ Space-Time Correspondence as a Contrastive Random Walk
Allan Jabri
PhD student, EECS, University of California, Berkeley
Abstract: This work proposes a simple self-supervised approach for learning a representation for visual correspondence from raw video. We cast correspondence as prediction of links in a space-time graph constructed from video. In this graph, the nodes are patches sampled from each frame, and nodes adjacent in time can share a directed edge. We learn a representation in which pairwise similarity defines transition probability of a random walk, so that long-range correspondence is computed as a walk along the graph. We optimize the representation to place high probability along paths of similarity. Targets for learning are formed without supervision, by cycle-consistency: the objective is to maximize the likelihood of returning to the initial node when walking along a graph constructed from a palindrome of frames. Thus, a single path-level constraint implicitly supervises chains of intermediate comparisons. When used as a similarity metric without adaptation, the learned representation outperforms the self-supervised state-of-the-art on label propagation tasks involving objects, semantic parts, and pose. Moreover, we demonstrate that a technique we call edge dropout, as well as self-supervised adaptation at test-time, further improve transfer for object-centric correspondence.
◦ Quantifying solidification of metallic alloys with scikit-image
Marianne Corvellec, scientific software developer, scikit-image core team and independent researcher, Institute for Globally Distributed Open Research and Education (IGDORE); and C. Gus Becker, PhD student, Center for Advanced Non-Ferrous Structural Alloys (CANFSA)
scikit-image is a well-established Python library boasting a wide collection of image processing algorithms. We present two applications of image processing to the field of materials science, where scikit-image allows for us to track and analyze solidification structures in metallic alloys as they evolve through time. The image data are time series of 2D x-radiographs obtained at Argonne National Laboratory (ANL) and Los Alamos National Laboratory (LANL). Understanding the solidification properties of metals pave the way to improve metallurgical processing and manufacturing techniques.
— Break (~14 mins)
Back to Program Summary
Day 2: TUESDAY, May 18
11:00 AM – 12:00 PM — Session 5: Health
◦ Session Introduction
Maryam Vareth
BIDS Health and Life Sciences Lead, and Co-Director, Innovate For Health, UCSF and University of California, Berkeley
◦ High throughput lesion evaluation and quality control for incorporating quantitative imaging metrics into clinical practice
Anisha Keshavan
Lead Data Scientist, Octave Bioscience
Abstract: Automated multiple sclerosis lesion counts and volumes are poised to be salient clinical biomarkers of disease progression; however, algorithmic variability and low expert agreement prevents widespread adoption in clinical practice. Because every method has a non-negligible error rate, visual quality control (QC) is required before a clinical decision can be made. QC is a bottleneck to the use of automated lesion count and volume metrics in the clinic. We developed a quick and scalable web application to simultaneously evaluate expert and non-expert raters, and QC lesions from an automated method. This enables us to 1) improve expert agreement on lesion identification, 2) develop better quality education materials for experts and non-experts alike, 3) train new raters quickly, and 4) ensure the quality of the image segmentation at scale.
◦ Pathology Image Analysis with Deep Learning
Saeed Hassanpour
Associate Professor of Biomedical Data Science, Computer Science, and Epidemiology; Geisel School of Medicine at Dartmouth
Abstract: With the recent expansions of whole-slide digital scanning, archiving, and high-throughput tissue banks, the field of digital pathology is primed to benefit significantly from AI and deep learning technology. This talk will cover new applications of deep learning for characterizing histologic patterns on high-resolution microscopy images for cancerous and precancerous lesions. Also, recent advances and future directions for developing and evaluating deep learning models for pathology image analysis will be discussed.
◦ Computational Image Phenotyping for Stroke and Prostate Cancer
Corey Arnold
Associate Professor, UCLA Departments of Radiology, Pathology, Bioengineering, and Electrical & Computer Engineering
Abstract: The UCLA Computational Diagnostics lab investigates data driven methods for extracting discriminative signals from healthcare data. Our work incorporates imaging, pathology, and clinical data in machine learning and deep learning frameworks, with an emphasis on multi-modality data fusion. This talk will discuss our work in using magnetic resonance imaging and digital pathology for stroke characterization and prostate cancer detection.
— Break (~10 mins)
Back to Program Summary
Day 2: TUESDAY, May 18
12:00 – 1:00 PM — Session 6: Tools & Methods II
◦ Session Introduction
Laura Waller
BIDS Faculty Council Member, and Associate Professor, Electrical Engineering and Computer Sciences, University of California, Berkeley
◦ The ImageJ Ecosystem
Kevin Eliceiri
Professor of Medical Physics and Biomedical Engineering, Morgridge Institute for Research and University of Wisconsin-Madison
◦ napari: a fast, interactive, multi-dimensional image viewer for python
Nick Sofroniew
Imaging Product Lead, Chan Zuckerberg Initiative
I will introduce napari, a fast, interactive, multi-dimensional image viewer for Python. napari is designed for browsing, annotating, and analyzing large multi-dimensional images. I will cover the basic principles and interface of napari, with an emphasis on how napari helps with visualization of large, out-of-core datasets. I will also discuss extending napari with plugins, and our broader vision for the napari plugin ecosystem.
◦ Automated quantification and analysis of human brain connections
Ariel Rokem
Research Assistant Professor, Department of Psychology and eScience Institute, University of Washington
Abstract: The white matter of the brain contains the long-range connections between distant cortical regions. The integration of brain activity through these connections is important for information processing and for brain health. Diffusion-weighted MRI (dMRI) provides estimates of the trajectories of these connections in the human brain in vivo and assesses their physical properties. In this talk, I will present a set of open-source software tools that automatically infer these trajectories and delineate major anatomical pathways, providing estimates of the tissue properties along their length. These tissue properties serve as the input to statistical models that learn the relationship between brain connection tissue properties and complex behavioral and cognitive phenotypes. Taking into account the grouping of brain features according to their anatomical structure, these models provide inferences that are both accurate and interpretable.
— Break (~10 mins)
Back to Program Summary
Day 2: TUESDAY, May 18
1:00 – 2:00 PM — ImageXD Community
For this social and networking hour, participants will be able to engage directly in conversations and discussions with speakers and other participants. Explore the lounges to find or to initiate a discussion that resonates with your interests and research.
BIDS ImageXD (Image Analysis Across Domains) is a cross-domain initiative that promotes interdisciplinary collaboration and training for researchers, scientists, and theorists interested in using graphs and network analysis for applications in a variety of fields across STEAM including (but not limited to) anthropology, art, biology, computer science, economics, history, linguistics, mathematics, physics, and sociology.