Deep learning excels in medical imaging segmentation but requires extensive annotated data. To reduce reliance on expert labeling, we propose a semi-automatic method, where a small subset of the available data is annotated to train a model. The model then generates pseudo-annotations for the remaining data, refining the model iteratively until no further improvement is observed. We demonstrate the method’s effectiveness by training models on the Kvasir-Seg and Breast Ultrasound Images Dataset using SegFormer and U-Net architectures. With only 30% of the annotated data, our method achieves at least 90% of the accuracy of models trained on the full dataset. We also propose a method to estimate optimal data annotation using the model’s entropy measure.