Data-driven Simulations For Training AI-Based Segmentation of Neutron Images

Description:

Neutron interferometry uniquely combines neutron imaging and scattering methods to enable characterization of multiple length scales from 1 nm to 10 μm. However, building, operating, and using such neutron imaging instruments poses constraints on the acquisition time and on the number of measured images per sample. Experiment time-constraints yield small quantities of measured images that are insufficient for automating image analyses using supervised artificial intelligence (AI) models. One approach alleviates this problem by supplementing annotated measured images with synthetic images. We disseminate synthetic and measured data for a set of nine known materials (calibration phantoms) that were imaged using a neutron interferometer acquiring four-dimensional images and segmented by AI models trained with synthetic and measured images and their masks.

Motivation:

This web-based visualization and interactive dissemination system has been designed to facilitate reusability and reproducibility of neutron beam imaging experiments targeting supervised image segmentation tasks. The motivation is to prepare a large number of pixel-level multi-class annotated images for training artifical intelligence (AI) based segmentation models.

Access to data downloads

• (1) Measured Intensities for 84 autocorrelation values: A sample with polystyrene (PS) beads with varying radii in Deuterium Oxide suspensions acquired at ORNL HFIR)
tar.gz file download (1.5 GB)
• (2) Manual Mask: 1 manually annotated mask with pixel-level annotations of 13 classes for the PS dataset (1)
tar.gz file download (52 KB)
• (3) Measured Intensities for 39 autocorrelation values: Amorphous Solid Dispersion (ASD) sample with varying concentration of Palmitic Acid-d31 in poly(lactic-co-glycolic acid) or PLGA acquired at ORNL HFIR
tar.gz file download (650 MB)
• (4) Manual Mask: 1 manually annotated mask with pixel-level annotations of 9 classes for the ASD dataset (3)
tar.gz file download (31 KB)
• (5) Simulated Train Intensities for 84 autocorrelation values: Derived from the PS dataset (1) for 15 checkerboard scene geometries per train+validation and per test evaluations
tgz file download (26.5 GB)
• (6) Simulated Test Intensities for 84 autocorrelation values: Derived from the PS dataset (1) for 15 checkerboard scene geometries per train+validation and per test evaluations
tgz file download (26.5 GB)
• (7) Manual Masks: 15 train + 15 test annotated checkerboard masks with pixel-level annotations of 13 classes for the datasets (5 and 6)
tar.gz file download (3.4 MB)
• (8) Simulated Train Intensities for 39 autocorrelation values: Derived from the ASD dataset (3) for 15 checkerboard scene geometries
tar.gz file download (10.6 GB)
• (9) Simulated Test Intensities for 39 autocorrelation values: Derived from the ASD dataset (3) for 15 checkerboard scene geometries
tar.gz file download (10.6 GB)
• (10) Manual Masks: 15 train + 15 test annotated checkerboard masks with pixel-level annotations of 9 classes for the datasets (8 and 9)
tar.gz file download (3 MB)
Note 1: Each intensity dataset consists of three modes: H0, H1, and Hdark.
Note 2: Simulated data (intensities and masks) are separated into two subsets, one for train+validation and one for test accuracy evaluations (i.e., altogether 30 simulated datasets).

Credit

If you use this data then, please, cite the appropriate publication:

Data-driven Simulations For Training AI-Based Segmentation of Neutron Images

Pushkar S. Sathe¹, Caitlyn M. Wolf², Youngju Kim³,⁴, Sarah M. Robinson³, M. Cyrus Daugherty³, Ryan P. Murphy², Jacob M. LaManna³, Michael G. Huber³, David L. Jacobson³, Paul A. Kienzle², Katie M. Weigandt², Nikolai N. Klimov³, Dan S. Hussey³, and Peter Bajcsy^1
1Information Technology Laboratory, NIST
²NIST Center for Neutron Research, NIST
³Physical Measurement Laboratory, NIST
⁴Department of Chemistry and Biochemistry, University of Maryland, College Park
(under review)