Why Image segmentation results not picking up small masks. How to improve the model for better detection

I implemented UNET architecture to detect hemorrhages on retinal fundus images. I was able to generate decent results but my model doesn’t seem to pick small masks and I was wondering if you can help to improve the scenario. Here are a few examples. (Left: Input RGB retina image, middle:actual mask, right:predicted mask) I circled some small masks.