[scientific research notes] about using the downsampling multiple of openslide cut

background

Medical images are huge , We call the whole picture WSI, That is to say whole slide image. We can't use such a large graph in actual processing , So we usually use small ones patch, these patch It is cut from the original picture .

For the convenience of narration , We assume that there is a (h, w) Big picture of .

level What is it?

When cutting a picture, we first involve a level The problem of . If it is level 0, So let's go straight to this (h, w) On the picture of . If it is level 1, So we are (h/downsample_factor, w/downsample_factor) This graph cuts up , In other words, the big picture was first reduced . If it is level 2, The size of the graph becomes [h/(downsample_factor^2), w/(downsample_factor^2)], That is to say, it is further reduced . That is to say, the graph will be adjusted to [h/(downsample_factor^level), w/(downsample_factor^level)].

Okay , that level and downsample_factor What is the relationship ？ We might as well look at . Reading medical images generally uses openslide This package , The code is as follows .

slide = openslide.open_slide(str(filename))

slide Is an instantiated object , Let's take a look at its properties slide.level_downsamples. This is actually a list , In this example , The list is [1.0, 4.0, 16.0, …], in other words , This is the difference level Down sampling multiples .level=0 When ,downsample_factor yes 1.0,level=1 When ,downsample_factor yes 4.0. Yes, of course , In different situations , The numbers in this list are not necessarily the same , We should make a concrete analysis according to the specific situation , And use slide.level_downsamples Check to confirm .

Tangent process

Take the first step and zoom in , Then cut the picture . We can set patch Size , such as 256, So all the pictures we cut out are 256*256. If we want to overlap, Then adjust the step size , For example, when there is no overlap, the step length and patch equally , Follow the example just now , That's it 256 Step size of . When it's all done , I want to have a look resize, That is to say, in the end, we should patch Zoom to how big .

preservation

Because there are coordinates in the picture name , So we have to know exactly what this coordinate is . In general , We can find the part of the coordinate definition .

    row_cords = np.arange(0, w, args.stride * downsample_factor)
    col_cords = np.arange(0, h, args.stride * downsample_factor)

It's not hard to see. , The coordinates are the original coordinates , Without any scaling , We should use appropriate coordinates according to our own needs .