Improving UNet segmentation of small objects

I have a previous model I’ve trained using Dlib to segment images that contain some very small objects, 1-3 pixels in size (It’s a binary segmentation problem). It got a pretty good Dice score (94.5) that I’ve been unable to replicate using Fastai (92).

My image size is 1344, and the main issue is with the very small objects which are getting missed. Does anyone have any experience/tips regarding this? I’ve tried a quite a few different things, but haven’t been able to match my Dlib model so far.

Most recently I’ve tried using different loss functions including Locasz, but it still didn’t help.

Thanks in advance!