Sensitivity Boundaries of Image Segmentation Models: Stress Tests of UniverSeg and IRIS on Non-Standard Images

Universal medical image segmentation models based on In-Context Learning (ICL), like UniverSeg and IRIS, which is In-context Reference Image guided Segmentation, can adjust to new tasks without many labeled datasets. However, the robustness boundaries of these models when dealing with non-standard images in real images, such as low-quality images or those with offset, remain unclear. The study uses a testing method called a data probing approach. Blur is added to the pictures or items are moved around a bit to see how well the models work. Research wants to see the performance of these models when the example pictures are not perfect, which is in daily use. The research shows that the performance of the models is related to the clarity and offset of examples. The UniverSeg model felt difficult when items moved more than 10 pixels and the Gaussian blur radius exceeded 10. The IRIS model is better at working with pictures but it has trouble when the pictures are moved, reaching 40 pixels. This study shows that the models used now are not perfect and need improvement. This report helps doctors and engineers to judge whether ICL models are supposed to be used in working situations.