microsoft/BiomedParse

What it solves

BiomedParse is a foundation model designed for comprehensive biomedical image analysis. It provides a unified approach to perform segmentation, detection, and recognition of biomedical objects across diverse imaging modalities, reducing the need for separate models for different medical imaging tasks.

How it works

The model uses a text-guided approach to identify and segment objects. In version 2, it incorporates the BoltzFormer architecture to improve the segmentation of small objects and supports end-to-end 3D volumetric inference. It can process 3D volumes slice-by-slice, encoding neighboring 3D context around each slice. It also includes a built-in object existence detection module to identify false positives without requiring separate mask checking.

Who it’s for

It is intended for researchers and practitioners in the biomedical field who need to analyze complex biomedical data across various modalities such as CT, MRI, Ultrasound, PET, and microscopy.

Highlights

• Multi-modality support: Works across nine modalities including CT, MRI, Ultrasound, PET, and various microscopy techniques.
• Broad anatomical coverage: Supports over 200 anatomies in v2.
• SOTA 3D performance: Achieves state-of-the-art performance in 3D biomedical image segmentation.
• Unified framework: Combines segmentation, detection, and recognition into a single tool.