Medical

Medical Image Segmentation

Segmenting organs and abnormalities in medical images.

4 datasets26 resultsView full task mapping →

Medical image segmentation delineates anatomical structures and pathologies in 3D medical scans — critical for treatment planning, tumor measurement, and surgical guidance. nnU-Net remains the dominant auto-configuring baseline, while SAM-Med2D and Med-SA bring foundation model capabilities to medical segmentation.

History

2015

U-Net (Ronneberger et al.) establishes the encoder-decoder architecture for biomedical segmentation

2017

V-Net extends U-Net to 3D volumetric segmentation for CT/MRI

2018

nnU-Net auto-configures preprocessing, architecture, and training for any medical segmentation task

2019

Attention U-Net adds attention gates for focusing on relevant anatomical regions

2021

UNETR applies vision transformers to 3D medical image segmentation

2021

nnU-Net wins multiple Medical Segmentation Decathlon challenges

2023

SAM (Segment Anything) adapted for medical images — SAM-Med2D and MedSAM

2024

STU-Net scales 3D medical segmentation to 1.4B parameters

2024

TotalSegmentator provides 117-class whole-body CT segmentation

2025

Foundation segmentation models enable zero-shot organ and tumor delineation

How Medical Image Segmentation Works

1Image Acquisition3D medical volumes (CT2PreprocessingnnU-Net-style preprocessing…3Encoder PathA CNN (ResNet blocks) or Vi…4Decoder PathThe decoder upsamples featu…5Segmentation OutputA pixel/voxel-wise classifi…Medical Image Segmentation Pipeline
1

Image Acquisition

3D medical volumes (CT, MRI, PET) or 2D images (X-ray, ultrasound, pathology slides) are acquired and standardized.

2

Preprocessing

nnU-Net-style preprocessing: intensity normalization, resampling to target spacing, and cropping to the region of interest.

3

Encoder Path

A CNN (ResNet blocks) or ViT encoder compresses the image into a hierarchy of feature maps at decreasing resolutions.

4

Decoder Path

The decoder upsamples features back to input resolution, with skip connections from the encoder preserving fine-grained spatial detail.

5

Segmentation Output

A pixel/voxel-wise classification produces a segmentation mask, with each pixel labeled as background or one of the target structures.

Current Landscape

Medical image segmentation in 2025 is dominated by nnU-Net — the auto-configuring framework that consistently achieves top results across diverse tasks without manual tuning. Foundation models (MedSAM, SAM-Med2D) are enabling interactive and few-shot segmentation, reducing the annotation burden. TotalSegmentator demonstrates that comprehensive whole-body segmentation is now production-ready. The field is mature for common structures (organs, major vessels) but still challenging for small lesions, rare pathologies, and domains with limited training data.

Key Challenges

Annotation cost — expert radiologist/pathologist annotation of 3D volumes is extremely expensive ($50-200 per scan)

Class imbalance — lesions and tumors are tiny relative to background, requiring focal loss or oversampling

3D memory constraints — processing full-resolution 3D volumes requires patch-based training with careful aggregation

Domain gaps — different scanners, protocols, and patient populations cause segmentation failures on new data

Evaluation metrics — Dice score alone is insufficient; surface distance and volumetric accuracy matter for clinical use

Quick Recommendations

Any new segmentation task

nnU-Net

Auto-configuring, consistently wins or ties SOTA on new datasets with zero manual tuning

Interactive segmentation

MedSAM / SAM-Med2D

Click-based prompting for quick annotation and novel structure segmentation

Whole-body CT

TotalSegmentator

117 anatomical structures in one pass, production-ready

Research with transformers

SwinUNETR / UNETR

Best transformer-based architectures for 3D medical volumes

What's Next

The frontier is universal medical segmentation — models that can segment any anatomical structure or pathology from any imaging modality with minimal or zero task-specific training. Expect foundation models pretrained on massive multi-institutional datasets, federated learning to access data across hospitals, and real-time segmentation for surgical guidance.

Benchmarks & SOTA

Related Tasks

Get notified when these results update

New models drop weekly. We track them so you don't have to.

Something wrong or missing?

Help keep Medical Image Segmentation benchmarks accurate. Report outdated results, missing benchmarks, or errors.

0/2000