**Contact Us**

CiiS Lab

Johns Hopkins University

112 Hackerman Hall

3400 N. Charles Street

Baltimore, MD 21218

Directions

**Lab Director**

Russell Taylor

127 Hackerman Hall

rht@cs.jhu.edu

**Last updated: 3/27/17 11:30 am**

Design, implement, and evaluate an algorithm that creates spatially dependent dose features at the inter-organ level to identify specific areas of the head and neck that are more or less critical and sensitive to radiation damage.

**Students:**Chris Micek, Arun Raghavan, Julie Shade**Mentor(s):**Dr. Todd McNutt, Dr. Russel Taylor

Radiation therapies are designed to minimize radiation exposure to areas outside the tumor to avoid adverse reactions to radiation therapy, such as swallowing impairment and speech damage. However, not much is known about which anatomical regions most increase or decrease the probability of specific adverse effects when exposed to radiation- there may be areas in the soft tissue that are absolutely critical to avoid. Currently, specific known structures can be avoided by contouring them semi-automatically from CT image data and planning radiation therapies to avoid them but it would take too much time to contour all structures in the complex head and neck anatomy. We plan to address this problem with the following solution:

1. Develop an algorithm to automatically define anatomical regions in the inter-organ space of the head and neck and determine the radiation dose to these regions.

2. Incorporate this software with machine learning algorithms to determine which of the specific areas in the head and neck are more or less critical and sensitive to radiation damage, to better assist in the planning of radiation therapy to minimize adverse effects.

**Minimum:**(Completed)- Well-documented API for creation of a statistical atlas of contoured anatomy and script to register a new patient to this atlas using coherent point drift. Script to predict the location of a defined cube-shaped region from the atlas in this patient, written in Python.

**Expected:**(Completed)- Well-documented expansion on methods from minimum deliverable to define the same arbitrary volume in multiple patients, corresponding to the same anatomical soft-tissue region. Statistical validation of this method showing minimal error between arbitrarily selected regions.
- Creation of multiple statistical atlases using different anatomical regions to determine which gives most accurate prediction of the location of the arbitrary volume in patients not included in the atlas.

**Maximum:**(Completed)- Previous: GUI to select anatomical region in statistical atlas and view ROI estimates along with dose volume histogram of carved regions in addition to the expected and minimum deliverables.
- Updated: Methods to map a 3D grid from the atlas coordinate space to the patient coordinate space and calculate the radiation dose in one tetrahedron at a time.

We first create a statistical atlas representing a point cloud of mean patient anatomy. We then define the desired anatomical region for dose-toxicity analysis in this patient space. Note that this can be any arbitrary region, as shown in red in Fig. 1, but will likely be an evenly spaced grid over some region of the patient space and dose-toxicity analysis will be performed for each section of the grid individually. Next, we query the Oncospace database, an in-house database of patient CT and dose data stored as run-length encoded binary masks, to find a point cloud of all anatomical regions that are contoured in the desired patient and included in the atlas. We use coherent point drift to register the atlas to the patient and fit a thin-plate spline to create a mapping for any (x, y, z) point in the atlas space to the patient space. We transform the desired region to the patient space using the thin-plate spline and calculate dose inside the desired region by querying the dose map in the Oncospace database. Each step of this workflow will be described in more detail and code can be found in the technical appendices. With this framework, our group will be able to use machine learning methods to determine whether radiation dose to any given anatomical region is predictive of adverse effects and avoid these regions in future radiotherapy plans.

**2a. Registration: Coherent Point Drift**
We chose to use a coherent point drift registration method to perform the necessary registrations. These include an iterative registration process to create the statistical atlas and registrations between the atlas and other patients, used to transform regions in the atlas coordinate system to the patient coordinate system. The method we used is described in the 2010 paper by Myronenko, A., & Song, X, Point Set Registration: Coherent Point Drift. This involves creation of a smooth vector field transformation between two point sets, which is ideal for our application involving anatomical structures. The detailed mathematical method can be found in their paper, but the pseudocode of the method is:

Initialize parameters: λ, β, σ

Construct G matrix, initialize Y = Yo

Deterministic annealing:

EM optimization, until convergence:

E-step: Compute P

M-step: Solve for W from:(diag(P · 1)G + λσ2I)W = PX - diag(P · 1) Yo

Update Y = Yo + GW

Anneal σ = α * σ

Compute the velocity field: v(z) = G(z, ·)W

**2b. Statistical Atlas Creation**
Our framework allows users to create an atlas representing a point cloud of mean patient anatomy using any number of patients from the Oncospace database and any number of anatomical regions. The documentation for code used to create this atlas is provided in the technical appendices. The procedure begins by user selection of a list of anatomical regions and number of patients to include. The database is queried for a set of point clouds representing the desired number of patients that have all the desired regions contoured. The centroids of each data set are aligned and a random patient’s coordinate system is chosen as the base coordinate system. All patient data sets are registered to this “fixed patient” coordinate system using CPD. Next, a statistical average of the data sets after registration is taken. The resulting average point cloud is set as the fixed patient coordinate system and the process is repeated until the registration error in below a given threshold value.

**2c. Transformation: Thin-Plate Splines**
To create a mapping function between the patient coordinate frame and the coordinate system of the atlas, we fit a Thin Plate Spline (TPS) transformation to the registration between the atlas and the patient. This is done by calculating the kernel matrix across the set of points given to the function. The kernel function Φ(z) for each point z consists of a 1 x K (where K is the number of control points chosen). From the kernel function a surface function is calculated and represented as a matrix. The pseudo-inverse of this matrix then can be used to transform any set of points from atlas coordinate space to patient coordinate space.

**2d. Dose Extraction**
A TPS transformation from atlas space from patient space allows us to find the dose in a patient of any region we might define in atlas space. We first define a dense, regular grid of points in the atlas with voxel-width spacing, and then use the computed TPS transformation to deform this grid onto the patient coordinate system. The dose information for each patient is defined in Oncospace as a coarse grid over a subset of the patient volume, with each “voxel” containing the dose administered in the corresponding area of the grid. Functions already exist to interpolate this dose information onto the binary mask representing the patient anatomy; we convert the deformed grid into a mask contained in the patient volume, and use these functions to get dose data for each voxel in the region of interest. Then it is simple to calculate all manner of statistics on the calculated dose, e.g. to obtain the maximum, minimum, and mean dose.

Access to anatomy database and pre-written python code for decompressing and displaying the data. (Complete, 2/19)

Install CRKit (Computational Radiology Kit, free from Harvard Medical). (Complete, 2/20)

Find optimal deformable surface registration algorithm. (Discuss with Dr. Taylor, Dr. McNutt, and their students, lit search; plan to complete by 3/3)

- Milestone name: Fully read and understand all code documentation. Complete UML diagram of expected additions to code.
- Planned Date: 3/3
- Completed Date: 3/4
- Status: Complete

- Milestone name: Working deformable registration method.
- Planned Date: 3/31
- Completed Date: 3/27
- Status: Complete.

- Milestone name: Statistical atlas creation method.
- Planned Date: 3/31
- Expected Date: 3/31
- Status: Complete.

- Milestone name: Validation of statistical atlas and point cloud transformation.
- Planned Date: 5/5
- Expected Date: 5/5
- Status: Complete

- Milestone name: Framework for automatic dose-region identification in regularly spaced grid.
- Planned Date: 5/18
- Expected Date: 5/18
- Status: Complete

See timeline below for milestones and status:

- Project Plan
- Project Background Reading
- See Bibliography below for links.

- Project Checkpoint
- Paper Seminar Presentation:
- Project Final Presentation
- Project Final Report
- Source code documentation
- Source repository (requires permission from group members)

G. Chintalapani and R. H. Taylor., “Construction of Multi-Component Statistical Shape Models of Bone Anatomy (Abstract in Proceedings)”, in Computer Assisted Orthopaedic Surgery (CAOS), 2009.

G. Chintalapani, R. Murphy, R. Armiger, J. Lepistos, Y. Otake, N. Sugano, R. H. Taylor, and M. Armand, “Statistical Atlas Based Extrapolation of CT Data for Planning Periacetabular Osteotomy”, in Medical Imaging 2010: Visualization, Image-Guided Procedures, and Modeling, 2010. pp. 539-548. DOI:10.1117/12.845570

G. Chintalapani, L. M. Ellingsen, O. Sadowsky, J. L. Prince, and R. H. Taylor, “Statistical Atlases of Bone Anatomy: Construction, Iterative Improvement and Validation”, in Proc. 10th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI), Brisbane, Australia, October, 2007. pp. 499-506.

Commowick, O., Grégoire, V., Malandain, G. (2008). Atlas-based delineation of lymph node levels in head and neck computed tomography images. Radiotherapy and Oncology, 87(2), 281-89. ISSN 0167-8140, http://dx.doi.org/10.1016/j.radonc.2008.01.018.

R. B. Grupp, H. Chiang, Y. Otake, R. J. Murphy, C. R. Gordon, M. Armand, and R. H. Taylor, “Smooth extrapolation of unknown anatomy via statistical shape models”, in Proc. SPIE 9415, Medical Imaging 2015: Image-Guided Procedures, Robotic Interventions, and Modeling, San Francisco, 8-10 Feb., 2015. p. 941524. 10.1117/12.2081310

R. Grupp, Y. Otake, R. Murphy, J. Parvizi, M. Armand, and R. Taylor, “Pelvis surface estimation from partial CT for computer-aided pelvic osteotomies,” in Computer Assisted Orthopaedic Surgery, Vancouver, June 17-19, 2015. p. (online).

Han, X. et al. (2008). Atlas-based auto-segmentation of head and neck CT images. In: Metaxas D., Axel L., Fichtinger G., Székely G. (eds) Medical Image Computing and Computer-Assisted Intervention – MICCAI 2008. MICCAI 2008. Lecture Notes in Computer Science, vol 5242. Springer, Berlin, Heidelberg

Han, X., Hibbard, L.S., O'Connell, N., and Willcut, V. (2009). Automatic Segmentation of Head and Neck CT Images by GPU-Accelerated Multi-atlas Fusion, MIDAS Journal. http://www.insight-journal.org/browse/publication/685

O. Sadowsky, G. Chintalapani, and R. H. Taylor, “Deformable 2D-3D Registration of the Pelvis with a Limited Field of View, Using Shape Statistics”, in MICCAI, Brisbane, Australia, 2007. pp. 519–526. PMID: 18044608

S. Seshamani, G. Chintalapani, and R. Taylor, “Iterative Refinement of Point Correspondences for 3D Statistical Shape Models”, in Medical Image Computing and Computer-Assisted Intervention (MICCAI), Toronto, September 18-22, 2011. pp. 417-425.

Teng, C., Shapiro, L. G., Kalet, I. (2006). Automatic segmentation of neck CT images. Proc. 19th IEEE International Symp. on Computer-Based Medical Systems (CBMS)

Warfield, S., Zou, K., Wells, W. (2004). Simultaneous truth and performance level estimation (STAPLE): An algorithm for the validation of image segmentation. IEEE Trans. Med. Imag. 23(7) 903–921

Zhang, T., Chi, Y., Meldolesi, E., Yan, D. (2007). Automatic delineation of on-line head-and-neck computed tomography images: Toward on-Line adaptive radiotherapy. International Journal of Radiation Oncology*Biology*Physics, 68(2), 522-30. ISSN 0360-3016, http://dx.doi.org/10.1016/j.ijrobp.2007.01.038.

Here give list of other project files (e.g., source code) associated with the project. If these are online give a link to an appropriate external repository or to uploaded media files under this name space.