Promptable Segmentation

What it does

Promptable Segmentation outlines structures on any medical image. Click inside a lesion, draw a box, scribble or lasso on a 2D image, a slice of a CT or MR volume, or a frame of an ultrasound or endoscopy video; the run returns a mask, and each further click refines it. The bare model ID runs this interactive tool.

From the mask it measures area or volume, and the mask propagation section carries a mask drawn on one frame or slice through the whole clip, cine loop or volume. Two related tools cover the cases without clicks: name a structure in text ("liver", "left kidney cyst", "pleural effusion"), or label every structure the modality supports in one pass.

Intended use

A general tool for annotating, measuring and correcting structures where no dedicated MedRun model applies, or where a dedicated model's mask needs a correction. A radiologist outlines a lesion with two clicks and reads its volume; a researcher segments every organ in a CT cohort; a developer propagates a mask through an ultrasound clip. A mask drawn from clicks carries the label the user gives it.

Who it is for

Developers and researchers building datasets and measurement pipelines, radiologists and other physicians who measure or correct structures, and teams producing volumetry or 3D models from CT and MR.

Accepted input

• Images: any series MedRun accepts, including CT, MR and PET volumes, ultrasound, angiography and endoscopy cine, radiographs and photographs; also NIfTI-1/2, NRRD, PNG, JPEG and MP4.
• Prompts: positive and negative points, boxes, scribbles or a lasso on a 2D image, a slice or a video frame.
• Reference mask (optional): an existing DICOM SEG or NIfTI mask to refine or propagate.

Requirements

Requirement	Why
Consistent slice spacing for 3D volumes	Volumes in millimetres and propagation between slices
Up to about 1,000 slices per call	Larger series are cropped to the body region of interest
Video of up to a few minutes	Propagation runs through every frame of the clip
At least one prompt for the interactive tool	The mask is defined by where the user clicks or draws

The modality is recognised from the series, and modality-specific normalisation is applied: Hounsfield windows for CT, intensity normalisation for MR, colour handling for ultrasound and endoscopy.

Optional context

A label for the structure ("right adrenal nodule") can be attached to the prompt, so the mask and its measurement are named in the result.

The result

Section	Content
Mask and measurements	The prompted mask, refined with each prompt, with area (2D) or volume (3D)
Mask propagation	A mask from one frame or slice carried through a video, cine loop or volume, as a multi-frame mask

The prompts are echoed back with the result, so a mask can be reproduced or refined later from the same clicks.

Standards

• DICOM: Segmentation Storage and Label Map Segmentation Storage with SNOMED CT segment category and type codes; RTSTRUCT on request; SR TID 1500 for volumes; GSPS for overlays on the source images.
• Other formats: NIfTI-1/2 masks for analysis pipelines; JSON with measurements[] and mask references.
• Terminology: SNOMED CT and RadLex for named structures.

Prompt tips

On X-ray angiography, prompt with clicks rather than boxes. Add a negative click to cut a leak into a neighbouring structure.

Related tools

• universal/segment/text: one mask per named target from a text prompt, with an existence score, in DICOM SEG with SNOMED CT codes.
• universal/segment/everything: automatic label map of the organs, bones and vessels known for the detected modality, with per-structure volumes in SR TID 1500.