How to find ground truth in 3 D volumetric brain tumor nii file.?

Ground truth of 3 D volumetric brain tumor nii file.?

Answer 1

[Walter Roberson]

You cannot do that.

Ground truth of a data set is an indication of what the data at that location really is. Not what some algorithm estimates it is: what it really is.

So for example, ground truth what species each plant in a patch of flowers is, can be determined by doing DNA analysis of each plant, since DNA is considered to define species.

Pictures of the plants involved might be able to help reduce the search, but pictures of plants cannot account for the fact that there are some species that can vary a lot from plant to plant depending on factors such as micronutrients and temperatures and amount of sunshine or which genes happened to get expressed.

Some species such as Apple vary a lot -- if you grow the seeds of an Apple tree, the tree you get out will not look much like the original. And two plants that look a lot alike are not necessarily the same species!

So you cannot count on regular visible-light pictures to distinguish whether a particular cell in a brain is part of a tumor or not.

Can you count on MRI or CT to distinguish whether a particular cell in a brain is part of a tumor or not? No, that turns out not to be reliable either.

The researchers I worked with found that the most accurate measurement they could come up with, was MRS, Magnetic Resonance Spectroscopy -- but you have to do a lot of studying to figure out what chemicals in what ratios indicate tumors. WIth MRS we were able to match or exceed the accuracies of brain tumor pathologists. But even so getting more than 90% accuracy was very difficult.

The typical .nii file simply does not have the data that would permit you to determine the ground truth of the images in the file. Sometimes the .nii is accompanied by another file that labels the tumor areas as determined by professionals... but keep in mind that the professionals tend to only be 84% to 87% accurate.

[hla hla myint]

Thanks Teacher

[Walter Roberson]

We actively worked on detection of cancer before visible changes existed. Chemical markers (such in urine) and MRS helped a lot. But the implications were that MRI or CT might look normal but the ground truth was that cancer was there.

We also worked on trying to automatically distinguish between harmless abnormalities and cancer. That turned out to be difficult. We found that in some cases what the algorithm noticed was some very subtle differences in the shadows at the edge. It wasn't necessarily the size or the brightness or the eccentricity of the patch but rather a small difficult-to-distinguish shading. Sometimes when I compared the images I could see what the algorithm was pointing out, but if I tried to use the knowledge to make predictions based on images, I was often wrong: the algorithm was picking up on patterns of twists. It was my belief that the algorithm was overtraining for those and that there was no real information content in the details it was indicating.

Now, automatically finding areas of interest that should be brought to attention was comparatively simple, but accurately deciding tumour or not was hard. And of course the ones that did not yet have visible presence would get overlooked.