in image segmentation problems, after your model outputs a prediction, try downsampling it (by 1/2 or 1/4) before upsampling it again. this could yield better results