Boundary Extraction Algorithm for Low-Resolution Ultrasound Images

Please refer to the sample ultrasound image we captured (Figure 1) showing a tumor resembling a doughnut shape. As you may know, ultrasound images suffer from low resolution if compared to CT scan images. Therefore, a very thin layer of tissue boundary is shown very thick (in ultrasound images) resulting in thick boundaries of targets when compared to higher-quality CT images.

For example, in Figure 2, we illustrate the expected boundary via orange-colored curves that looks like a doughnt. If the same target was imaged with a CT scanner, we should see only a boundary of hair-line thickness. This has been shown in Figure 3 (CT will show accurately this red-line as the boundary).

Given the diverse shapes these tumors can exhibit, how can we develop an algorithm to accurately extract their boundaries from surrounding tissue or blood in ultrasound images?

To clarify, in cases like the doughnut-shaped tumor presented, without prior knowledge, the algorithm might incorrectly identify a coin-shaped target.

Any insights on how to approach this challenge would be greatly appreciated.

Answer 1

[Image Analyst]

I see no justification from the image alone that the bright region(s) would form a donut shape. You must be indicating those using apriori expert knowledge that the tumor is a perfect donut shape, perhaps because it's an image of a known phantom, right?

My suggestion is to denoise the original grayscale image (not the pseudocolored image),

1. blurring it with imfilter or using a median filter medfilt2,
2. then threshold it,
3. take the convex hull with bwconvhull,
4. use bwboundaries on the convex hull,
5. then fit the (x,y) data to a circle using the attached function.

That will at least get you a guess at the outer circle.

[Sardius]

Thank you very much for your response and for providing prompt feedback. I acknowledge the need for improvement in image quality. I also concede that our device, currently in prototype stage, produces images of lower quality compared to existing commercial devices. However, when considering images captured by commercial devices, the same query persists. Please refer to Figure 1 below. While the ring is visible to the naked eye, extracting and isolating it from the image proves challenging due to other areas of similar intensity, etc. How can we accurately determine boundaries in such images?

Regarding the code (circlefit.m), it's worth noting that the shape of the targets may vary, and it's not always a ring or circle.

[Image Analyst]

I suggest you look here for image processing algorithm dealing with ultrasound:

http://www.visionbib.com/bibliography/contentsmedical.html#Medical%20Applications,%20CAT,%20MRI,%20Ultrasound,%20Heart%20Models,%20Brain%20Models

There are any number of weird shapes in that/those image(s) and I think it's only because of your expert knowledge that you are able to delineate something in the image that's not completely there.

https://www.sciencedaily.com/releases/2014/05/140506115622.htm

https://en.wikipedia.org/wiki/Pareidolia

I know that the thing you're seeing is really physically there, but not completely shown in the image, but your eye can sort of "fill in the gaps" and see more or less the complete object. If thresholding what's actually there does not delineate the parts of the thing that are not actually shown in the image then you might try deep learning

https://www.mathworks.com/help/vision/ug/semantic-segmentation-using-deep-learning.html

or try activecontour to do Chan Vese segmentation.

https://www.google.com/search?hl=En&as_sitesearch=www.visionbib.com%2F&bthG=Google%2BSearch&q=chan+vese#ip=1

or

https://www.google.com/search?q=chan+vese+segmentation&sca_esv=ab02dd2696e7e4b9&sca_upv=1&sxsrf=ACQVn09Bi-aUYh8GqCMLRxcKrowpiJK4fw%3A1711487692139&source=hp&ei=zDoDZqGUBp3dptQP9IWb4Ac&iflsig=ANes7DEAAAAAZgNI3DGfkeEw5wcoeSVwmqVGy8H5JQK-&gs_ssp=eJzj4tVP1zc0LCkxzSsvTykxYPQSS85IzFMoSy1OVShOTc9NzStJLMnMzwMAAT4NvA&oq=chan+vese+&gs_lp=Egdnd3Mtd2l6IgpjaGFuIHZlc2UgKgIIAjIEECMYJzIKEAAYgAQYigUYQzIKEC4YgAQYigUYQzIFEAAYgAQyBRAAGIAEMgUQABiABDIKEAAYgAQYigUYQzIGEAAYFhgeMgYQABgWGB4yCBAAGBYYHhgKSLU-UABY7BBwAHgAkAEAmAHJAaABygqqAQUyLjcuMbgBAcgBAPgBAZgCCqACqQvCAgoQIxiABBiKBRgnwgITEC4YgAQYigUYxwEYrwEYjgUYJ8ICChAuGIAEGIoFGCfCAhcQLhiABBiKBRiRAhixAxiDARjHARjRA8ICCxAAGIAEGIoFGJECwgIIEAAYgAQYsQPCAhEQLhiABBixAxiDARjHARjRA8ICERAuGIAEGIoFGJECGMcBGNEDwgIOEAAYgAQYigUYsQMYgwHCAgsQABiABBixAxiDAcICFhAuGBQYgwEYxwEYhwIYsQMY0QMYgATCAgsQLhiABBixAxiDAcICDhAuGIAEGLEDGIMBGNQCwgIKEAAYgAQYFBiHAsICBRAuGIAEwgIHEAAYgAQYCpgDAJIHBTEuOC4xoAftbA&sclient=gws-wiz