2
$\begingroup$

Soon I will be working on biomedical image segmentation (microscopy images). There will be a small amount of data (a few dozens at best).

Is there a neural network, that can compete with U-Net, in this case?

I've spent the last few hours searching through scientific articles that are dealing with this topic, but haven't found a clear answer and I would like to know what other possibilities are. The best answers I found are that I could consider using ResU-Net (R2U-Net), SegNet, X-Net and backing techniques (article).

Any ideas (with evidence, not necessarily)?

$\endgroup$
0
$\begingroup$

Hey i am working on my Bachelor thesis at the moment and use UNET in combination with a GAN for image segmentation. I spend the last 5 months on that, so on my tests, the new approach of januar 2020, called Multires-UNET is quite a good choice for more texture orientated segmentation. I use the current github implementation. Its quite nice, maybe you notice that you can easyly tweak the number of parameters with "alpha" in the implementation, to scale the multiple Resnets in the Unet structure.

I tried also other segmentation networks like Mask_RCNN with different backbones or tried to construct various types of CAE on my own, but always had to come back to a UNET like structure. Same goes for ResU-NET and R2U, the multires one worked better for my purposes cause i didnt need any kind of LSTM modules.

Some examples which may clarify the difference in performance on a specific task:

Original Image: enter image description here

Ground-Truth: enter image description here

Classical UNET++ (Unet with skip-connections)(2.5 million parameters) more parameters (wide) didnt change the result. enter image description here

Multires-UNET (alpha=1.67, think was about 7 million parameters) enter image description here

Can you show some of ur microscopy images, how complex is ur task and what do you want to segment ?

| improve this answer | |
$\endgroup$
  • $\begingroup$ Hi there, may I ask, what does "more textured orientated segmentation" mean? How do you decide that a picture you are performing segmentation on is more textured orientated? :) Will certainly check it out. I can't show the images right now, after this coronavirus-situation I will get my hands on those $\endgroup$ – Nuwanda Apr 4 at 13:08
  • $\begingroup$ Hope the images give you some insigth, i trained it on about 500 handmade GT (used ImageJ for preprocessing, very useful tbh). Works way better for colored images, but Grayscale is a hard test for texture identification. $\endgroup$ – Paul Higazi Apr 4 at 13:29
  • $\begingroup$ With texture orientated i basicly meant that it searches for deeper or more color/rotation/graysclae - INVARIANT features. So features which you cant fool that easy with some augmentation for example like color shift. $\endgroup$ – Paul Higazi Apr 4 at 13:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.