on the left we have binary masks that highlight only a flat, geometric shape with corresponding physical metadata communicating with the computational "pixel resolution" that because of the current imaging hardware constrains, produces a minute distortion from reality known as anisotropy due to artificial STITCHING.
leaning right end we might have neural networks like TCNs that capture every single pixel on the image with no binary distinction of each pixel's representation, and treat every brightfield object's internal, at most blurry aspects, and its outer peripheral environments somewhat fairly. with a bit of noise.
on the very right end, the constraint of data complexity is eliminated by the hardware capabilities that intake images from all angles that capture the entire 3 dimension of the object, thus representing its structure entirely. this has much higher noise.
beyond that spectrum (further right) are destructive, such as IHC or flow cytometry to find scRNA-seq data. they have the highest level of noise.
the key is to find the right balance to understand how much meaningful features we need from the data itself, reduce the model from be exposed to too much noise to ensure it knows its own limitations and bounds of knowledge gaining. we tell it "here is all you need to know, and the reason why we hide the rest is because they will distract you, because their nature is more stochastic."