SlicerMorphCloud: An interactive, cloud-based platform for quantitative analysis of high-resolution 3D scans of biological specimens.

Ali Maga, University of Washington

ACCESS Allocation Request BIO180006

Abstract: Many core biological sciences courses such as Evolution, Ecology, Natural History and others rely on biological specimens as teaching aids. COVID pandemic fundamentally changed how these courses are taught. Instructors had to replace hands-on laboratories that used physical specimens with virtual laboratories, thanks to availability of 3D digitized natural history specimens available on aggregate repositories like MorphoSource (https://morphosource.org). As 3D specimens becomes more widely available, there is a corresponding need for image analysis software that supports open, accessible, and reproducible research. Our lab have been developing SlicerMorph (https://github.com/SlicerMorph/SlicerMorph) since 2019 as a comprehensive extension to the 3D Slicer, an open-source biomedical image processing platform (https://slicer.org) . SlicerMorph extends the existing core capabilities of Slicer (such as 3D rendering, segmentation and quantification) with modules to better support non-clinical imaging datasets and facilitate analysis of 3D organismal form through geometric morphometric methods (aka statistical shape analyses). SlicerMorph provides convenience modules for data retrieval, import and export from high-resolution research microCT scanners. Our team supplements these software development activities with intense short-courses, workshops to train and support the biosciences community along with self-paced tutorials (https://github.com/SlicerMorph/Tutorials). All these activities are funded from a number of NSF/BIO grants (ABI-1759883, ABI-1759637, ABI-1759839). While SlicerMorph makes working with 3D specimens simpler and accessible to all biologists, there are still certain roadblocks. Particularly most biology students and labs that are not computationally oriented do not have the necessary computing resources required to process these datasets. E.g., a modest 2 gigavoxel (1024x1024x2048) 3D microCT scan of a fish can require up to 48GB of memory depending on the data type representation (short vs float) and the requirements of the imaging filter (e.g., Grow from the seeds) being applied. Medium size datasets (4-8 gigavoxels) may consume 80GB, while larger datasets (10+ gigavoxels) may require upwards of 100GB of RAM transiently. With pandemic this became particularly challenging for courses which try to offer virtual laboratories that rely on these data. To fill this need gap, we have developed the SlicerMorphCloud platform in 2021 that leveraged JetStream (now JetStream2) cloud farm to deploy remote, interactive desktop sessions using docker containers. When we started, we have two specific uses cases for SlicerMorphCloud: 1. A backend to support the biannual online short-courses in SlicerMorph 3D Image Analysis and Morphometrics. Each of these short-courses run for a week and average between 50-65 attendees. Setup docker containers with all the tutorials and sample data ready to be run by each attendee. This greatly simplified the logistics of offering an interactive computational workshop, increased the efficiency of the workshops. So far, we have done 4 online workshops with over 220 attendees using this approach. 2. To support our users by providing short-term access to powerful virtual desktop environment for biologists to help facilitate their data collection and analysis efforts. In both use cases, users login to the SlicerMorphCloud via a browser session (noVNC and related technologies) and immediately start utilizing the platform without needing to install any software. With the new GPU and high-memory instances offered by the new JetStream2 (JS2) platform, we plan to supplement these two uses cases with two new ones: 3. Train and deploy deep-learning based segmentation models for the community: One of the most time-consuming tasks in digital morphology is the interactive segmentation of anatomical structures so that their volumes, shapes and other properties can be quantified. We plan to use the powerful A100 GPUs to train segmentation models and deploy them for the community by leveraging the open source Medical Open Network for Artificial Intelligence (MONAI). MONAI is a freely available, community-supported, PyTorch-based framework for deep learning in 3D imaging. MONAI also includes MONAILabel, a client/server based, open-source image labeling and learning tool that helps researchers collaborate, create annotated datasets, and build AI models in a standardized MONAI paradigm within Slicer (https://github.com/Project-MONAI/MONAILabel). 4. Open-source stereophotogrammety pipelines: As it stands, the majority of the 3D biological specimen data has been generated by microCT imaging. However, this has been changing thanks to the emergence of powerful stereo-photogrammety pipelines. In a nutshell, stereophotogrammety uses large number of 2D images acquired from a specimen in different orientations to build a 3D model by first finding corresponding features in each picture and then triangulating the 3D shape based on these features. Commonly called Structure From the Motion (SFM), these pipelines can be used to generate 3D models of large landscapes (e.g., forest habitats from drone captures), as well as small objects (e.g., rodent skulls) in lab settings. Primary benefits of this approach is the ability to provide texture information in addition to the 3D geometry, which can be valuable for different types analysis (e.g., habitat classification in ecology) and the rather reduced cost of data acquisition. However, SFM pipelines are computationally expensive, as they require hundreds to thousands of pictures to build accurate models. Because of the importance of this type of data to a broad number of biological disciplines, we plan to deploy open-source OpenDroneMap (ODM) software on the JS2 nodes for the community use. We have already successfully tested deploying a web-based prototype (https://smc-1.slicermorph.org:8000) with a small number of users. It should be noted that we are submitting an NSF/BIO proposal to the Infrastructure Capacity for Biological Research solicitation to build a portal to consolidate these tools into a single interface for easier long-term maintenance and convenience of our user base.

Allocations:

2022 Indiana Jetstream2 Storage 20,000.0 GB
2022 Indiana Jetstream2 Large Memory 1,269,760.0 SUs
2022 Indiana Jetstream2 GPU 1,121,280.0 SUs
2022 Indiana Jetstream2 710,080.0 SUs
The estimated value of these awarded resources is $395,163.20. The allocation of these resources represents a considerable investment by the NSF in advanced computing infrastructure for the U.S. The dollar value of the allocation is estimated from the NSF awards supporting the allocated resources.
2021 Indiana Jetstream2 Storage 10,000.0 GB
2021 Indiana Jetstream2 GPU 600,000.0 SUs
2021 Indiana Jetstream2 600,000.0 SUs
The estimated value of these awarded resources is $138,500.00. The allocation of these resources represents a considerable investment by the NSF in advanced computing infrastructure for the U.S. The dollar value of the allocation is estimated from the NSF awards supporting the allocated resources.
Click to show/hide prior allocations »
2018 IU/TACC (Jetstream) 740,240.0 SUs
2018 IU/TACC Storage (Jetstream Storage) 2,048.0 GB
2018 The Science Gateways Community Institute (SGCI) Yes
The estimated value of these awarded resources is $14,907.20. The allocation of these resources represents a considerable investment by the NSF in advanced computing infrastructure for the U.S. The dollar value of the allocation is estimated from the NSF awards supporting the allocated resources.

Other Titles:

Click to show/hide prior titles »
3D Quantitative Phenotyping Gateway (3DQPG): A Community Resource for Quantitative Phenotyping of Biological Structure