Constellation (CON-001)

Constellation (CON-001) is an automated radiological image processing software that utilizes whole-body MRI scans from healthy adults to provide quantitative measurements of various anatomical structures. Its methodology combines non-AI algorithms and convolutional neural networks for image segmentation, offering clinicians a decision support tool through detailed PDF reports with visualizations and volumetric data.

Constellation (CON-001) is an automated image post-processing software designed for clinical MRI settings to assist physicians. It processes whole-body DICOM MR images acquired from a GE MR450W machine, specifically for healthy adult patients. The software's methodology involves combining images from overlapping anatomical stations, then utilizing a hybrid approach of non-AI based algorithms (for brain, lungs, femur, lower limb muscles, subcutaneous adipose tissue) and convolutional neural networks (for kidneys, spleen, liver, visceral adipose tissue) to segment various anatomical regions. These regions include lower limb muscles, visceral and subcutaneous adipose tissue, kidneys, liver, spleen, lungs, and brain structures. The system performs atlas-based registration for brain and leg muscle segmentation and is compatible with T1 MPRAGE, 2-point DIXON, and 6-point DIXON sequences acquired in axial or sagittal planes. The primary output of Constellation is a PDF report containing quantitative measurements such as organ, muscle, and fat volumes, alongside whole-body visualizations with alpha-blended segmented label images. This report serves as a clinical decision support tool and is not intended for automated diagnosis, triage, emergency medicine, or critical care; clinicians retain final diagnostic responsibility. Performance testing validated its efficacy, with segmentation accuracy for various organs and tissues evaluated using metrics like the Dice Similarity Coefficient (DSC) and mean percent absolute difference. Brain segmentation accuracy was assessed via mean percent absolute difference and Pearson’s correlation. The system also underwent device repeatability, test-retest measurement repeatability, software verification, and usability testing, demonstrating performance according to its intended use.

AI Regulatory Concerns

Powered by

Jev reads the full FDA letter and returns calibrated concern probabilities per regulatory category. These are AI estimates, not regulatory advice.

• Q1. What is the specific performance (e.g., average Dice Similarity Coefficient) of the convolutional neural network models for segmenting kidneys, spleen, liver, and visceral adipose tissue compared to the non-AI algorithms used for other structures?; How does the system handle variability in MR image quality or artifacts from the GE MR450W scanner, and what are the specific image quality threshold checks implemented?; For which specific "healthy adult patient" populations was the system validated, and are there any known limitations regarding age, BMI, or other demographic factors within this group?; The document mentions "atlas-based registration for segmentation of brain regions and leg muscles." What atlases are used, and how is their applicability ensured across diverse healthy adult populations?; How does Constellation's whole-body quantitative measurement accuracy compare to other established methods or gold standards for assessing body composition and organ volumes beyond the listed predicate devices?