33 Computer Vision News Medical image segmentation has long been a specialized field, requiring tailored models for different imaging modalities and datasets. However, a novel approach said to be the first foundation model for promptable medical image segmentation promises to change the landscape. The model in question, Segment Anything in Medical Images – MedSAM for short – is unique in its ability to handle diverse medical imaging modalities accurately and efficiently, providing a one-size-fits-all solution that can significantly improve a physician’s workflow efficiency. “Before MedSAM, most medical image segmentation was a specialist approach where, for different modalities and datasets, you had to train different models,” Bo tells us. “This was for two reasons. One, we didn’t have much data. Now, we’re at a tipping point where we can access millions of medical images. Two was the model architecture. MedSAM inherited from the Segment Anything Model (SAM) published by Meta Research. We took that base model and continuously fine-tuned SAM with millions of medical image masks.” MedSAM
RkJQdWJsaXNoZXIy NTc3NzU=