Joey-S-Liu/MedSAM3

What it solves

MedSAM3 is designed to simplify medical image segmentation by removing the need for manual prompts like bounding boxes or points. Instead, it allows users to segment medical targets using pure text-guided medical concepts (e.g., "skin lesion").

How it works

The project uses a parameter-efficient fine-tuning strategy based on the SAM3 model. It employs LoRA (Low-Rank Adaptation) to adapt the base model to a large-scale, diverse medical dataset containing over 658,000 images and nearly 2.9 million instance annotations across 330 unique medical concepts.

Who it’s for

Medical imaging researchers and clinicians who need a flexible, robust segmentation tool that works across various modalities, including radiology (CT, MRI, PET, X-ray), optical/microscopic imaging, and video-based procedures (ultrasound, endoscopy, surgery).

Highlights

• Text-Guided Segmentation: Segments targets based on medical concepts rather than manual coordinates.
• Broad Modality Support: Covers radiology, microscopy, histopathology, and surgical videos.
• Bespoke Medical Dataset: Trained on a massive dataset of 658,094 images and 2,863,974 annotations.
• Efficient Tuning: Utilizes LoRA for parameter-efficient adaptation of the base SAM3 model.