Why do lower-resolution images sometimes yield better results in deep learning for bioimaging analyses? 🤔📉 Mariana Ferreira's new preprint on #ReScale4DL explores this paradox and introduces optimal resolution design! @gomez-mariscal.bsky.social brainchild🧑‍🔬✨ Check: www.biorxiv.org/content/10.1... - ThreadSky

henriqueslab.bsky.social • 47 days ago

Why do lower-resolution images sometimes yield better results in deep learning for bioimaging analyses? 🤔📉

Mariana Ferreira's new preprint on #ReScale4DL explores this paradox and introduces optimal resolution design! @gomez-mariscal.bsky.social brainchild🧑‍🔬✨

Check: https://www.biorxiv.org/content/10.1101/2025.04.09.647871v1.full.pdf

Comments

ivankashkan.bsky.social•46 days ago

Cool. Maybe a diletant question, but what happens if you train a model on a set of images of different resolution (exactly same as one you generate for finding ideal image resolution/receptive field match)? Will it work as a kind of regularisation?

ivankashkan.bsky.social•46 days ago

*augmentation*, not regularisation.

truenorth-ia.bsky.social•26 days ago

Q1 Did you do any comparisons where the receptive field is kept constant (or as constant as possible) in physical units? This would increase computational requirements and for some problems not be worth it, but should address the field of view issues seen with high resolution images.

truenorth-ia.bsky.social•26 days ago

Q2 Did you use the tensorflow version of your semantic segmentation notebooks? Any plans to move to Pytorch? (I know you guys set up a lot of things on the cloud, but for Windows Users the tensor flow (and stardist) installation is becoming increasingly convoluted.

truenorth-ia.bsky.social•26 days ago

(for Q1 when I say 'keep receptive field constant in physical units' I mean increase it in pixels for the high resolution images, so it covers the same physical space... increasing problem size of course, which may or may not be worth it depending on the problem).

gomez-mariscal.bsky.social•25 days ago

Thank you very much for the questions!🤩 Q1: Exactly, the accuracy reported is for the same FOVs (with different dimensions). In supp fig (microSAM), we only show the differences for a colocalised patch. To keep the FOV, one should apply some tiling with the results of those individual patches.

gomez-mariscal.bsky.social•25 days ago

Q2: This work is independent of TF or PyTorch. We choose publicly available and user-friendly networks, but the pipeline can be replicated with any other model. Indeed, microSAM is developed with PyTorch

guijacquemet.bsky.social•47 days ago

Congratulations 🎉 🎉 🎉

docdellaire.bsky.social•47 days ago

Important! Also might save a lot of terabytes if lower bit depth patch sizes yield good results. Will read and pass on.

senger.bsky.social•47 days ago

this is super interesting

Comments

Posting Rules

Reply