Transformational Technologies for
Biology, Medicine, and Health
Honors 222B & MEBI 498A, Spring, 2013
Project #2:
MRI image data, segmentation and labeling
Due May 2, noon
In this assignment, you will gain experience doing some basic image analysis and manipulation, as you work with a Magnetic Resonance Imaging (MRI) data set. This is an individual project, but see below about helping classmates. The assignment should also teach you about the challenge of automatic image segmentation, as well as something about the wealth of detail inside human anatomy. Please keep in mind that auto-segmentation is an unsolved (and possibly unsolvable?) research problem in medical image analysis. Thus, this is a open-ended assignment, with the goal being simply to learn as much as you can in a week about image processing, medical segmentation and anatomy.
Steps for this project: (You need not follow this exact order)
- First, look at the visible human browser from NIH and the University of Michigan. (I'll demonstrate this in class). From this site, download two selected images (use high resolution) that you'll use for your analysis below.
- Next, look at the MRI data set (with axial and sagittal views) of the lumbar vertebra. Again, pick two images for analysis, one axial and one sagittal. (Note that you can choose to upload a single large file, 'AllImages.zip', if you want to browse all 48 images).
- Download and install the ImageJ program (public domain software from the NIH).
- Using ImageJ, do your best to try and find shapes, isolating them from noise and other entities in your four selected images. To do this, first crop your images to the area of interest. Then, try out a variety of operation. At a minimum, you must apply "find edges", and then test out various orderings of smoothing and sharpening operators. In class, I will provide a brief tutorial on some relevant imageJ functions.
- Finally, to complete the segmentation task, you need to label the objects. You do not need to hand-contour the objects; hopefully step 4 provided at least some places where the objects are reasonably segmented. You do need names for the anatomical entities. The cannonical souces of anatomic names is the the Foundational Model of Anatomy and you can use this browser to hunt around. However, a much better source is be the labels provided by the "thoracic viscera" atlas at http://www9.biostr.washington.edu/da.html. Of course, Wikipedia is also a good source of names and images; e.g. look at the "intervertebral disc" wikipedia page.
Collaboration:
In the past, some students have been frustrated with either (a) ImageJ bugs/failures, or (b) not knowing where to start or what things to try. Therefore, although this is an individually-graded assignment, and although you should all pick different images, I'd like to encourage a form of collaboration via a Catalyst dicussion board. (Email is okay too, but it is nicer and more effective if all of you see all of the comments / questions / ideas.) You can post to this discussion board if you are stuck, you can post the Image J operations that work well (or not), you can point your peers to other resources you may have discovered etc. This is obviously an experiment, but my hope is that the discussion board will encourage a level of brainstorming that isn't possible with a solo project.
What is required, per academic honesty and good ethical behavior, is that your deliverables give credit where credit is due, naming your peers who helped along the way. See below.
Deliverables:
In the class dropbox for project #2, I'll want at least eight images and notes from each of you. This must be handed in as a SINGLE FILE. You may either assemble a Word document, or use a powerpoint file. The images are large-ish, and I don't want you to reduce them, so expect a file size of 5-10 Meg or more.
- Provide the four initial images you started from (with filenames or visible human image numbers), as well as four segmented (or at least processed) images.
- List the manipulation and processing steps you took in imageJ (including parameter settings) to get from each of the starting images to each of the (partially) segmented images. If you include intermediate, or "dead end" images, include the processing steps for these too. These should be like lab notes -- detailed enough for your to reproduce the process.
- In each of the segmented images, tag the FMA entity names (or anatomic names) onto the outlines of the objects you segmented in the images. (You could do this easily in a powerpoint file, for example.) You should try to find a total of 10-20 anatomic entities across your images. See also the grading rubric about "learning anatomy".
- Include notes (not a formal essay) about the process. You may want to describe what was easy and what was not within ImageJ, as well as how accessible and usable the FMA explorer was for tagging the entities in the images. These notes should be embedded within the same powerpoint or Word document, and please make it is clear which images you may be referring to in the notes (i.e., number or title the images.) You may optionally also include intermediate images, if you think these are interesting or make an interesting point about the process. You should conclude with some rough impression about the whole process -- what you learned or what was interesting.
- If approapriate, add a section for acknowledgements. This is where you must give credit to your peers for ideas & suggestions that you tried out or used in your project.
Grading rubric:
Unlike project #1, this assignment does not include a formal essay; however, your "notes" must still be clear, and must include some summary statements about the overall process. The two most important factors for evaluation will be (a) evidence that you explored and learned about some "slice" of anatomy (via the FME and other resources) and (b) evidence that you explored and tested a variety of ImageJ processing choices for image manipulation, especially segmentation. Where appropriate, I'll also be looking for correctness and clarity as usual. Note that I will NOT be grading based on how well you truly segment the image--instead I'll just look for a reasonable exploration of ImageJ functions, and a reasonable understanding of the segmentation problem. See below for my grading rubric for this project:
2.5 -- 2.9 |
3.0 -- 3.3 |
3.4 - 3.7 |
3.8 -- 4.0 |
|
| Anatomy | Student does not sufficiently label images. | Student demonstrates only minimal exploration of anatomic parts. | Student has learned a significant portion of anatomy; sufficient entities are identified and labeled. | Student shows a wide variety of anatomic entities, and uses FMA names for labeling. |
| ImageJ processing & steps | Student uses only a minimal set of ImageJ operators. | Student has explored at least a few other ImageJ processing options beyond "find edges" and a smoothing operator. | Student provides evidence of extensive exploration of ImageJ capabilities. Student describes dead-ends attempts, as well as different sorts of successful segmentations. | Exemplary segmentation is demonstrated; notes include explanations of why different operators work well. Student researched some advanced operators. |
| Notes | Summary is lacking or incomplete; notes may have some clarity problems; I cannot easily understand when the student applied which operators to which images. | Notes are clearly written and sufficient for me to replicate what was carried out for each image. A solid summary is included with overall impressions about the imaging / segmentation process. |
Notes include insights about why certain operators may or may not have worked well with certain images. Summary includes implications for the broader field of imaging informatics. |
Student explores additional readings about image segmentation. |
Last Updated:
April, 2013
Contact the instructor at: gennari@u.washington.edu