My project title is Retinal vessel segmentation, I have completed two steps.
First step is construct direction map.
[0 0 0 1 0 0 0;
0 0 0 1 0 0 0;
0 0 0 1 0 0 0;
0 0 0 1 0 0 0;
0 0 0 1 0 0 0;
0 0 0 1 0 0 0;
0 0 0 1 0 0 0 ]
This is one of the mask (T_d)(there are 12 masks like this). Green channel of a image is taken for the process.
T_di is centered on each pixel p of G and the following parameters are measured:
The arithmetic mean, Adi(p), of the gray-levels of pixels of G included by T_di and aligned along di.
The arithmetic mean, NAdi(p), of the gray-levels of all pixels of G included by T_di and non aligned along di.
The difference, diffi(p) = Adi(p)- NAdi(p) Once all inspections have been accomplished, to assign the direc- tion to each pixel p of G the following parameters are computed, for k = 1,2, … , 12:
m(p) = mink {diffk(p)}
M(p) = maxk {diffk(p)}
This process is for identify the direction of a pixel. If M(p) > 0.25×(12-m(p)), the pixel p is considered as tentatively belonging to a vessel and is assigned in DM the value i representing the direction di at which diffi (p) has the highest value. Otherwise, no direction is considered as the most reliable one for p. The pixel is interpreted as belonging to the background and is assigned in DM the value 0.
Second step is edit direction map
i want to create a matrix Aux from DM as follows
Aux pixels are set to the gray-level 255 before performing the process for any given direction. The value 255,white, is chosen since it identifies the gray-level of pixels certainly belonging to the background. During the ith inspection, only and all the pixels of DM with value in the range [i -2, i + 2] are copied onto Aux, where are assigned value 0. In fact, the pixels currently copied onto Aux were (tentatively) interpreted as vessel pixels during the construction of DM, and it is then reasonable to assign to themthe gray-level 0, black,which is definitely a value for foreground pixels.
Condition checking is we remark that for any i (i = 1,2, … , 12), i + 1and i + 2 are computed modulo 12; as for i -1and i - 2, if these values result to be ≤0 they are increased by 12, so as to avoid copying in Aux pixels with i = 0, which does not correspond to a valid direction for foreground pixels. Again using AUX i created Direction map same as the first step.
My problem is i got output for both the steps, but in edited direction map pixels are scattered very much. I attached the figure file.
Can anyone help me to explain how to tune that pixels?
Hi Kalai,
That is quite the question! I believe I submitted a possible solution to measuring local directionality of your retina vessels on another post of your's using MIPAR (http://MIPAR.us). Did this not end up being satisfactory? I have attached it below.
Also, please view https://youtu.be/E1oSf8cVA5I to see what we can do fairly easily to segment retina vessels.
Cheers,
John
Dear Kalai,
in image processing the most important step is preprocessing! Aside from the impressive job of MIPAR and comparing it to YOUR first results, I'd suggest to extract the green channel and find appropriate low-pass filters (there is a lot of them) to get rid of the artefacts (noise, other structures) BEFORE applying the directional masks.
This step may of may not remove some of the finest vessels (difficult to assess in the color images) but will surely remove splotches and so on.
AND: fixed thresholds may or may not be applicable for your problem. Reviewing the results of every step will help to decide whether your thresholds are ok or need some adjustments.
Good luck!
Dreher sir,
I have a step to detect small blobs in an image, so in this work there is no preprocessing,in the result of edited direction map the vessel pixels are dispersed. I dont know what is the problem and how to overcome it? Can you help me?
Dear Kalai,
looking into your output showed artifacts I first misjudged as blobs. Having only the full-colored image 21_training to compare it was a bit difficult to compare.
BTW: the marked location is not the only location where your algorithm runs into problems. I'll attach a processed image - derived from 21_training(-3).tif to show what I can see.
Firstly, what did I do ?
1. Set the red and blue channels to zero.
2. Convert the green channel to gray.
Now, what do we see comparing the grayscale image to your output? Wherever the contrast is low (I'd suspect these vessels lie slightly deeper in the tissue), your algorithm starts to fail.
Obviously because your settings assume a contrast higher than available in some regions. Thus you may either discard these regions or adapt your parameters to the contrast available. One method might be to improve contrast locally, depending on the neighborhood. Increasing the contrast for the image globally might not be appropriate.
Hope this helps.
Hi Kalai, as I understand it, your problem is that in the resulting image, there are many isolated small groups of white pixels left ('scattered pixels'). The easiest way to get rid of these pixels is to remove all objects smaller than a specific threshold (e.g. 10 px). In Matlab, you can use bwareaopen(), see the example in http://de.mathworks.com/help/images/ref/bwareaopen.html
- Badges
- Science topic