Dear Wadha,

I am a histologist. I ask you please to develop your question. For example, What exactly you mean by "registration to a medical image"?

Yes, could you elaborate on what you want to do?

Histology images (in my experience) have much more pixels per centimeter. But that should not be a problem for registration: the registration method should not assume equidistant pixels in the two modalities. Registration speed is faster when down-sampling. Noise is usually low in histology images.

A problem is to get a good starting point for the registration. Keeping track of where and in what orientation the (stack of) histology slices are made, is not easy. Another problem is that tissue deformations can be huge, so the transformation of for example MRI to histology is not linear. You need deformable registration.

Having mentioned some problems, I want to add that such (for example post-mortem) registration has great value: in cancer research it provides the only evidence of tumor extent. So, important work!!

Dear Wadha,

Great question!

We do not need to always down-sample images for image processing anymore.

Down-sampling used to be the case in previous decades due to ease of computations, and just like you said for filtering and denoising. Thats not the case anymore. One can propose and do image processing techniques without any downsampling.

Best regards,

J.

Hi Marcos and Lambert,

what I meant is using a registration algorithms to wrap or geometrically align multi-modal images, for example, warping a histology image to MRI image). For this purpose, practitioners try to work with low resolution histology image which requires down-sampling the histology.

The question is: why do practitioners down sample the histology image to register it to (i.e MRI ) ?.

Marcos Celio De Almeida Lambert Zijp Marcos Celio De Almeida

Hi Wadha and all,

One of the most important parameters for a histological photography (digital or not) is resolution and sharpness. Blurring on the side may be the correlate of noise.

I will mention an almost daily procedure we do: photograph using a camera, acquiring the image in tif or other format. Normally 300dpi. To publish we need 600 dpi , generally. Normally I use Adobe photoshop for that. The image become more sharp, but the resolution is not improved ( it is a characteristic of acquisition instrument, microscope). One other characteristic, it acquired in this process is that it become computationally heavy. This last aspect could explain why it is not easy to work computationally with that. Working with low sharpness images is fast, however you lost quality! The obvious way to work is experiment using low sharpness images and after find the better sequence to register work with high sharpness and slowly! A very high quality image at the acquisition is very fundamental.

Marcos Celio De Almeida , let's discuss this some more.

The 300dpi or 600dpi you mentioned are printing resolutions I assume. I don't understand why the 600 dpi becomes sharper. Less pixelated, and more pleasing to the eye (assuming photoshop uses a good interpolator), but not sharper. Interpolation does not sharpen images, on the contrary!

I agree a stack of histology images needs a lot of memory and could become computationally heavy. By downsizing that can be solved, and the noise becomes less too, but is noise a problem? I think the much greater resolution of histology images, compared to CT or MRI (typical pixelsize is 1 mm), is the reason.

How many MB is a 'typical' histology slice anyway? Pixelsize in the order of 10 micron?

Wadha moh Alyami , could you upload a couple of dicom histology images?

Dear Lambert,

I think we have probably to separe microscope resolution: capacity to resolve two points as distinct . The less the distance between points higher the resolution.

Printing resolution does not add nothing to this microscope resolution. I doubles the pixels per area from 300 to 600. I describe a sensation I have when I do that.

If we think that adding pixels not related to real image increase noise than I think you are right. The interpolation done sensorially appears however not increase blurring. How do you define sharpness of a image?

Hi Marcos,

Sharpness is best evaluated in the frequency domain. The higher the frequencies (in the spatial domain: the steeper the gradients), the sharper the image. Interpolation adds pixels whose values are usually between the original values. The frequencies then become less, so the sharpness decreases. I say 'usually', because theoretically one could fit higher order polynomials to interpolate, but that comes with a risk of introducing high bias (i.e. adding 'features' that are not real).

In other words, interpolation often is more appealing to the eye (less pixelated), but is a kind of blurring. Necessarily so, in order not to introduce bias.

Hi Lambert,

I agree. I think that a related concept

is contrast, as sharpness can only be achieved if we have bondaries with high contrast.