search

JJGO / UniverSeg

UniverSeg: Universal Medical Image Segmentation
Apache License 2.0
504 stars 51 forks source link

OCTA image segmentation result is not good #20

Closed zhouziyiaichifeiniu closed 8 months ago

zhouziyiaichifeiniu commented 9 months ago

Thanks for your kind words. So, the paper talks about UniverseG being pretrained on the OCTA500 dataset. But when I tried it on my own OCTA dataset, things didn't quite match up. And here's the kicker - when I cranked up the image batch-size to 50, the results were way worse than when it was less than 5. Any idea why that might be happening?

adalca commented 9 months ago

Could you show some results (inputs, outputs, context-sets) ? In what intensity range are your images?