Удосконалення процесу анотування медичних зображень для задач сегментації за допомогою текстових підказок

Abstract

This work investigates the potential of the Language Segment-Anything model for extracting regions of interest (ROI) from medical images. By integrating text prompts into the Segment Anything Model (SAM) framework from Meta and developing a user-friendly application using Streamlit, this work demonstrates how such an approach enables the extraction of pseudomasks of regions of interest without the need for prior data annotation. This simplifies the image segmentation process, making it faster, more accurate, and more accessible for users. The proficiency of the Language Segment-Anything model in handling natural language instructions and its ability to reduce the need for manual data labeling holds significant promise for the field of medical imaging.