Hey team, hoping for some help here.
The images look great, but I'm running into a roadblock for quantification of neurites in hippocampus, any suggestions?
Details: I have several coronal rat brain slices (50um thick) which have three abs on them (primaries added as a cocktail, anti- TH, p-75, and ChAT) HRP- conjugated secondaries added individually and washed (anti- mouse, goat, and rabbit) each with an amplification step (TSA --> FITC, TRITC, and Coumarin) [these are super old animals]. I used a Nikon A1 scanning confocal to get ~.8um sections (from 11-20ish sections each) at 20x. In order to make sure results are quantifyable, the exact same microscope settings are used to capture each image stack.
Ive used ImageJ to split color and then manually adjust brightness and contrast of each stack such that adjusted levels match one another for each color. (this has been optimized to a "one size fits most" level). This is the step I believe may be causing problems, but let me continue.
This is followed by subtracting the background with a standardized "rolling ball" algorithm. Finally, the image is made binary using the auto-threshold function “Moments” (Tsai’s method attempts to preserve the moments of the original image in the threshold result; Tsai, 1985). From here the ROI boundary is thrown on and "analyze particles" is used to quantify objects by %area.
The problem is that when I use these "universal" brightness/contrast levels across all stacks of the same color (in the same region of the hippocampus) I get one result that looks good, but seems to hide a substantial number of neurites that may be important. However when I try to include the multitude of "missing signal" by creating images based on "relative levels" (in which each stack gets a custom brightness/contrast modification), the results are strikingly different. While I trust the universal levels to be comparable in regards to absolute levels of antigen present (necessary for results) I feel like there is a large amount of information that is missing (number of antigen-positive axons). I have come to the general conclusion that the %area obtained from the "universal" or "absolute" levels is a reflection of just that, the absolute levels of antigen being detected (of course this doesn't account for linear ranges above the detection threshold...) and that by modulating the brightness/contrast individually for each stack I am better able to detect number of antigen-positive structures.
Am I right to think this? Is there a better way to go about accurately assessing actual levels of these antigens? (we'll be doing Westerns too, but the cytoarchitectural information is critical to understanding potential mechanisms of action).
For the image, note the WT is on top, and TG is on bottom (greyscale is the z-stack), absolute levels on the left and "relative" levels on the right.
Dear Anthoni.
As you suspect, one of the issues most likely stems from your manual adjustment of the brightness and contrast, especially if your adjustments cause some pixels to be saturated or get clipped at zero.
If you have acquired your images with the same microscope settings, then using an auto threshold like Moments SHOULD yield the same area regardless or any linear adjustments you do on the brightness and contrast. The fact that this is changing suggests that you are saturating some pixels after the brightness and contrast adjustment, which modifies the distribution of the intensity histogram upon which Moments tries to calculate a threshold.
There are more things that are going on that you should inform us of if you want further help.
1. Are you systematic in the depth of acquisition? Depending on your sample, a stack acquired at 30 um from the coverslip can have lower intensities than one acquired at 10 um from the coverslip due to aberrations caused by refractive index mismatches or tissue opaqueness.
.2 You have stacks. What do you do with them? Do you run the analysis on the stacks or on a projection?
3.What kind of background are you subtracting with the rolling ball? Seeing your mages, you have conditions where the structures of interest lie closer to each other, causing the rolling ball algorithm to probably subtract less background than for a condition where the structures are more separated. It might be wiser to subtract or divide the image by a heavily gaussian blurred version of the stack, to make sure that you are treating the background the same way, more independently of the size and proximity of your structures.
What happens if you do not go for a 'manual' brightness and contrast adjustment?
Seeing that your structures are tube-like I would suggest applying a tubeness filter that enhances structures of the size you have here and apply an auto-threshold on these. This might give a less biased result in terms of areas.
However without access to the original data, it will be hard to give further advice.
Best
Olivier, thank you for your response!
In regards to your questions,
1. I have been acquiring the stacks such that I capture as many individual sections as possible from top to bottom usually providing about 17 sections that have obvious signal and are clear. 11 was the smallest stack and so, to minimize refractive aberrations, I decided to take the 11 middle-most sections from each stack to process further.
2. Analyze particles is run on the stack, and separately on a collapsed z-stack as a means of variation control. The %area from the z-stack is similarly hit-or-miss in regards to a strong signal:noise ratio. The z-stack is created following the rolling-ball background subtraction, prior to "make binary".
3. I'm really not sure what "background" is being subtracted. (full disclosure, my laboratory is primarily electrophysiology so I've had limited resources in regards to IHC image processing education. I like your idea regarding the subtraction/division of a heavily Gaussian-blurred stack. Do you have any syntax for ImageJ that would point me in the right direction for this?
Changes in, or lack thereof, brightness/contrast change the structures/pixels that pass the threshold for detection when made binary and eventually counted. Without some manual adjustment, large swathes of antibody-positive structures are uncounted, or the background is counted.
I've tried playing with a couple built-in macros for detecting "tubeness" in the stacks with little to no luck in producing repeatable and valid results. I have attached the original file for the stack shown as the "WT" in my above image.
Do you have any further suggestions?
I very-much appreciate your aid!
Best,
- Badges
- Science method