Everything Everywhere All at Once... with a single prompt

Following the direction set in one of our previous papers, we’re excited to share that our new article, “Prompt Once, Segment Everything: Leveraging SAM 2’s Potential for Infinite Medical Image Segmentation with a Single Prompt,” has just been published. Both papers are driven by a common question: Given the enormous resources typically required to develop specialized artificial intelligence models for a single task, could foundation models be used instead—without the need for costly optimization? To explore this, we focus on a task that is both technically challenging and critically important: medical image segmentation, where accurate results can directly impact patient care and save lives.

In this paper, we introduce a novel approach that harnesses the video segmentation capabilities of SAM 2 to reduce the number of prompts needed to segment an entire volume of medical images. The study begins by comparing the performance of SAM and SAM 2 in medical image segmentation. We then present our new method, which builds on SAM 2’s strengths. The results demonstrate that SAM 2 achieves an average improvement of 1.76 % in Jaccard Index and 1.49 % in Dice Score over SAM.

Our next objective is to develop a methodology that helps define the limits of what can be achieved using these foundation models.

We’re sure we’re missing something—after all, we can’t be everywhere at once. 😉 Got ideas? Come and see us, and let’s talk.

Prompt Once, Segment Everything is one of the outcomes of musicgenia, a project funded by the MICIU/AEI (grant CPP2021-008491, DOI: 10.13039/50100011033) and supported by the European Union through NextGenerationEU/PRTR.