U.S. flag An official website of the United States government

U.S. Food and Drug Administration

RAMAC: Registration based Automated Matching and Correspondence

Catalog of Regulatory Science Tools to Help Assess New Medical Devices 

Registration-based Automated Matching and Correspondence (RAMAC) is a tool that automatically identifies corresponding locations of landmarks across multiple images.

Technical Description

Registration-based Automated Matching and Correspondence (RAMAC) [1] is a software tool that automatically identifies corresponding locations of landmarks across multiple images. For example, in a longitudinal imaging study, multiple lesions may be annotated across many DICOM image series. This tool is used to find annotations in all image series that correspond to the same lesion.  The inputs to RAMAC are a set of images, along with coordinates for the landmarks of interest in each image. The algorithms in RAMAC performs image registration and landmark matching. The output is the correspondence of the landmarks associated with the same lesion across the image series.

The input images must first be converted to SimpleITK format [2]. The input landmarks of interest are in physical coordinates. In a preprocessing stage, users can choose to clamp the values in the images to a fixed range. This is useful in the analysis of abdominal CT images, in which air pockets may generate spurious gradients that hinder image the registration workflow. Next, image registration is performed on the input images, with one designated as the fixed image (Figure 1, Image A) and the other the moving image (Image B). Currently, RAMAC supports rigid registration using a stochastic gradient descent optimizer and linear interpolation. The registration workflow generates a rigid transform () that is applied to the landmark coordinates in the moving image. The transformed coordinates are then matched with the landmark coordinates in the fixed image using an augmented Hungarian method [3]. There may be cases in which a corresponding landmark does not exist in one of the images. In this case, RAMAC assigns a value of None to indicate the missing landmark. The correspondence can be outputted as a .csv file. The region-of-interests (ROIs) of the corresponding landmarks can also be plotted with RAMAC for visual inspection of the results.

RAMAC

Figure 1: Overview of RAMAC. The inputs to RAMAC are a set of images, along with coordinates for the landmarks of interest in each image. The algorithms in RAMAC performs image registration and landmark matching. The output is the correspondence of the landmarks in each image.

Testing

RAMAC was tested using numerical phantoms. Synthetic, spherical lesions with prespecified centroid coordinates were inserted into a 3D Shepp-Logan phantom [4] (Image A). The phantom and lesion coordinates then translated and rotated in 3D to simulate a moving image (Image B). RAMAC was then used to register Image B to Image A and recover the corresponding lesions. We showed the RAMAC could achieve its intended purpose in this test case. A Jupyter notebook illustrating this test case is distributed with the package. Additionally, the pipeline was tested on a CT dataset to identify corresponding metastatic lesions with ground truth correspondences generated by visual inspection. This method had been presented at the 2024 IEEE International Conference on Prognostics and Health Management (IEEE PHM) and reported in Ref. 1 below. 

Limitations

  • RAMAC relies on image registration to identify corresponding landmarks. If two images have no overlap of field of view (e.g., if the images are pelvic CT and chest CT), then the algorithm will fail to find correspondence.
  • Image registration accuracy depends on many factors, including image quality and patient position, deformation, and motion. If two landmarks are very close to each other, RAMAC may fail to correctly identify correspondence.

Supporting Documentation

The software documentation is provided at: https://github.com/DIDSR/RAMAC. The user manual and documentation is available at: https://ramac.readthedocs.io/en/latest/. 

[1] S. Mukherjee et al., “Image registration based automated lesion correspondence pipeline for longitudinal CT data,” 2024 IEEE International Conference on Prognostics and Health Management (ICPHM), Spokane, WA, USA, 2024, pp. 184-192, doi: 10.1109/ICPHM61352.2024.10627649

[2] B. C. Lowekamp, D. T. Chen, L. Ibáñez, and D. Blezek, “The Design of SimpleITK,” Front. Neuroinform., vol. 7, 2013, doi: 10.3389/fninf.2013.00045.

[3] H. W. Kuhn, “The Hungarian method for the assignment problem,” Naval Research Logistics, vol. 2, no. 1–2, pp. 83–97, Mar. 1955, doi: 10.1002/nav.3800020109.

[4] L. A. Shepp and B. F. Logan, “The Fourier reconstruction of a head section,” IEEE Trans. Nucl. Sci., vol. 21, no. 3, pp. 21–43, Jun. 1974, doi: 10.1109/TNS.1974.6499235.

Contact

Tool Reference