FocalPlane features… reproducibility in imaging – recordings

For the October edition of FocalPlane features…, we hosted a webinar on reproducibility in imaging with talks from Helena Jambor and Kota Miura. Helena told us, ‘How not to lie with image data’ and focussed on data presentation. Helena also touched on the importance of accessibility when presenting data and much of the discussion following the talk was about accessible LUTs and the use of grayscale images. In his talk, Kota discussed how to ensure your bioimage analysis is reproducible and why this is so important. A key take home message was the importance of recording exactly how you have performed your analysis including software versions. Of course, this information should be easily accessible within your publication.

You can access the summary of Helena’s talks here: https://hackmd.io/@HelenaJambor/BkqHPK01kg

Kota’s slides are available here.

We had a few questions that we didn’t manage to get to following the talks, you can find Kota’s answers below:

Q: If one is doing ratiometric analysis and segmentation channel is also bit variable e.g. DAPI or H2B for nuclei, is it stil ok because we are reporting relative values? I have problem understanding relative vs absolute signal intensity in these terms. 
It should be OK as the same segmented region is used for the numerator and denominator. Whether the ratio can be compared between different regions e.g. cells, depends on the careful examination of the difference in the segmented area with dark and bright cells.   

Q: What about proprietary software like Harmony which does not disclose the exact algorithms used – should these be allowed in publications? 
Ideally, completely black-box tools are better not to be used for scientific measurement, but if you can calibrate that tool and linear response is expected in the range of your measurement values, it should be fine. 

Q: Is there another way to do thresholding using intensity measurements that does not include using 2 channels? I am afraid that if I stain for 2 different proteins in the same sample using 2 dyes each, that would be 4 channels and there would be a chance of bleed-through… 
One way is to use constant area and shape surrounding each target object for measuring the intensity. For example, in the case of nuclei, find the largest nucleus and define a rectangular bounding ROI. Use the same-sized ROI to measure the intensity of other nuclei. Don’t forget to subtract the baseline intensity. 

Q: I’m wondering if AI-based segmentation might be a better choice if you are restricted for channels? It seems that something like stardist does a good job detecting objects with different intensities. 
Delineation performance is often scored against human annotation, so what matters then is how precise the human annotation is in predicting the real boundaries of object. Detection (e.g. counting nuclei, tracking nuclei) is another issue, which is less affected by the precision of boundary delineation. For this purpose, Stardist can be incredibly useful.  

Q: Great Talk! Regarding the first thresholding issue, what is your option for the ML/DL method for single-channel image segmentation?  
Almost the same answer as above. I believe that segmentation models for ML should be trained also with molecular signatures, which should perform better compared to human annotations.   

Q: If one wants to segment gfp tagged proteins which form clusters, to predict their area. In this case marking the circumference of these protein clusters is not a choice. Then how should we measure the area of protein clusters using intensity to segment? 
Calibrate the intensity to estimate the density of GFP. Define the lowest density which you would call “Cluster,” and use that value to determine the threshold value that separates the cluster from the background. This is also a type of segmentation by intensity thresholding, but an important step is that YOU decide the threshold value based on some scientific reason, not by the analysis of the shape of the histogram.  

(1 votes, average: 1.00 out of 1)