| name | medical-imaging |
| description | Medical image analysis guidance for BiomedParse segmentation results. Use when interpreting 2D or 3D medical image segmentation, discussing imaging modalities, or explaining segmentation outputs. |
Medical Imaging Analysis
Guide for interpreting BiomedParse medical image segmentation results
and discussing medical imaging concepts.
Supported Modalities
BiomedParse v2 supports these imaging modalities:
- CT (Computed Tomography) — lung, abdominal, cardiac
- MRI (Magnetic Resonance Imaging) — brain, cardiac, musculoskeletal
- X-Ray — chest, bone
- Ultrasound — abdominal, cardiac, obstetric
- Pathology — histological slides
- Endoscopy — GI tract
- Dermoscopy — skin lesions
- Fundus — retinal imaging
- OCT (Optical Coherence Tomography) — retinal layers
Analysis Types
2D Analysis
- Input: PNG/JPG images
- Output: Segmentation masks + heatmaps (NPZ format)
- Use case: Single-slice analysis, pathology, dermoscopy
3D Analysis
- Input: NIfTI volumes (.nii.gz)
- Output: 3D segmentation masks + heatmaps (NPZ format)
- Use case: Volumetric CT/MRI analysis, organ segmentation
Interpreting Results
When discussing segmentation results:
- Masks: Binary regions identifying the segmented structure
- Heatmaps: Probability maps showing model confidence (0.0-1.0)
- Threshold: Default 0.5; lower values capture more area with less certainty
Prompt Design for Segmentation
Effective prompts for BiomedParse follow anatomical terminology:
- Organ-level: "heart", "liver", "kidney", "lung"
- Structure-level: "left ventricle", "aorta", "portal vein"
- Pathology: "tumor", "nodule", "lesion", "hemorrhage"
Limitations
- Results are for research purposes only, not clinical diagnosis
- Model confidence varies by modality and anatomical region
- 3D analysis requires GPU for reasonable performance
- Always correlate with clinical findings and expert review
