AI models for Medical image segmentation

MedSAM2

Vendor: Bo Wang Lab, University of Toronto

What it does: segments any structure you point at in CT, MRI and ultrasound, including through a 3D volume, with one click or box per structure. No training needed, which makes it the fastest route to a working pipeline.

nnU-Net v2

Vendor: German Cancer Research Center

What it does: configures and trains itself to your own annotated cases and remains the accuracy reference for organ and tumour segmentation. The right choice where you hold annotations and accuracy governs.

TotalSegmentator v2

Vendor: University Hospital Basel

What it does: labels 117 anatomical structures in a whole-body CT in a single pass, with no prompt and no training. Used for automatic measurement across an entire archive.

SAM-Med3D

Vendor: Shanghai AI Laboratory

What it does: works natively in three dimensions rather than slice by slice, so a structure stays consistent through the volume instead of drifting between slices.

STU-Net

Vendor: Shanghai AI Laboratory

What it does: pretrained on large CT collections and fitted to a new structure from a small number of annotated cases, which suits structures no released model covers.

MONAI Auto3DSeg

Vendor: NVIDIA and the MONAI consortium

What it does: selects, trains and ensembles segmentation models for your dataset automatically. Heavier to run, and the option when the target has no suitable model at all.

MedSAM

Vendor: University of Toronto (Bo Wang lab)

What it does: outlines any structure you box on a 2D medical image, across ten modalities, with no training at all — the quickest way to get usable masks out of a mixed archive.

Median Dice 0.85–0.92 across 10 imaging modalities with box prompts (independent, Nature Communications 2024).

LiteMedSAM

Vendor: University of Toronto (Bo Wang lab)

What it does: does the same job as MedSAM at a fraction of the compute, which is what makes it viable on every image rather than on a sample.

Dataset-dependent; optimised for much faster inference than MedSAM.

SAM-Med2D

Vendor: OpenGVLab

What it does: segments from a click or a box, trained on 4.6 million medical images and nearly 20 million masks — the largest prompted 2D medical training set released.

Dataset-dependent; trained on 4.6M medical images and 19.7M masks, evaluated with Dice and IoU.

SegVol

Vendor: SegVol authors / BAAI

What it does: segments a 3D volume from a phrase, a box or a point, and needs far fewer prompts than slice-by-slice approaches because it works in three dimensions natively.

Competitive 3D Dice with far fewer prompts than slice-wise SAM (developer-reported).

MedLSAM

Vendor: OpenMedLab

What it does: finds the target in a 3D volume and then segments it, so an operator does not have to locate the structure first.

Dataset-dependent; reported with mean IoU for localisation and Dice for segmentation.

SAT

Vendor: SAT authors

What it does: segments from a text prompt across a very large vocabulary of structures, which suits work where the target list is long and changes.

Task-dependent across large-vocabulary medical segmentation datasets.

BiomedParse-v2

Vendor: Microsoft Research

What it does: segments a 3D volume from a phrase across CT, MRI, ultrasound, PET and microscopy, and also reports whether the object is present at all — first place in the CVPR 2025 challenge for this task.

Task-dependent; first place in the CVPR 2025 text-guided 3D biomedical segmentation challenge according to the official repository.

BiomedParse-v1

Vendor: Microsoft Research

What it does: segments, detects and labels more than a hundred biomedical object types from text across nine modalities, on 2D images.

Task-dependent across 100+ biomedical object tasks and nine modalities.

VISTA3D / VISTA-3D / NV-Segment-CT

Vendor: NVIDIA / Project MONAI

What it does: segments 127 CT classes automatically and lets a clinician correct the result interactively, which is usually what gets a mask signed off rather than re-drawn.

Dice about 0.87 across 127 CT classes; interactive correction supported (developer-reported).

VISTA-2D

Vendor: NVIDIA / Project MONAI

What it does: segments 2D medical and microscopy images, as the 2D counterpart of the same toolchain.

Task-dependent; reported per dataset.

NV-Segment-CTMR

Vendor: NVIDIA / Project MONAI

What it does: covers more than 300 anatomical classes across CT and MRI with no prompt, which makes whole-archive measurement a scheduled job rather than a project.

Task-dependent across 300+ supported classes; no single universal Dice score.

TotalSegmentator MRI

Vendor: University Hospital Basel

What it does: brings whole-body segmentation to MRI across 80 structures on any sequence, and beat other public tools by a clear margin in independent evaluation.

Dice 0.839 on the internal MRI test set across 80 structures; beat other public tools 0.862 against 0.759 on 40 structures (independent, Radiology 2025).

MONAI Model Zoo bundles

Vendor: NVIDIA / Project MONAI

What it does: supplies ready-trained bundles for specific targets — spleen, pancreas, lung nodule, multi-organ, whole-body CT — so a common task does not need a training run.

Bundle-specific, typically Dice 0.85–0.96 on the source challenge dataset (developer-reported per bundle).

Universal Model

Vendor: Universal Model authors

What it does: segments abdominal organs and tumours, optionally directed by a language target, from a single CT volume.

Task-dependent across abdominal organ and tumour classes.

SegmentAnyBone

Vendor: SegmentAnyBone authors

What it does: segments bone in MRI across anatomical sites, which general soft-tissue models handle poorly.

Task-dependent across anatomical sites; evaluated with Dice.