Research map / 3D vision / 3D generation
Towards a Comprehensive, Efficient and Promptable Anatomic Structure Segmentation Model Using 3D Whole-Body CT Scans
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CT-SAM3D, a promptable 3D SAM-style segmentation model trained on the curated TotalSeg++ whole-body CT dataset (107 anatomies), evaluated for zero-shot interactive segmentation against SAM-family and automatic methods.
Problem
Adapting SAM to 3D medical volumes is inefficient (2D adaptation slices volumes) and needs many prompt clicks; existing 3D SAM methods lack comprehensive whole-body anatomic coverage and efficiency.
Approach
A 3D promptable segmentation network (ResT-based encoder, 64^3 patches) with a light-weight CPP prediction network that uses overlapping sampled patches, plus a curated whole-body CT dataset adding skeletal muscle, visceral fat, and subcutaneous fat annotations to TotalSeg.
Key results
- On FLARE22, CT-SAM3D reaches 88.4 mDSC and 95.8 mNSD, above nnU-Net (83.6 / 88.5) and TotalSeg Model (86.5 / 94.1).
- CT-SAM3D uses 77.30M params, 125.70 GFLOPs, 0.14 s per 64^3 patch and 1240 MB GPU memory, lower than SAM-family baselines in memory.
- Reported time per organ for CT-SAM3D correction is under 10 s versus about 0.5-2 h for manual editing.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use interactive CT-SAM3D instead of nnU-Net | FLARE22 segmentation · FLARE22 | mDSC: 83.6 -> 88.4 | Helped |
Surprising findings
- 3DSAM-adapter results hardly improve with more clicks, suggesting dataset-specific fine-tuning erodes interactive ability.
Limitations
- Unstable performance on extremely incomplete structures with fewer than 100 annotated voxels (e.g., Vertebrae_C6, Humerus_R).
- Single-granularity annotations; finer substructures such as Couinaud liver segments need extra interactions.
- The system cannot yet automatically extract semantic information of anatomic structures.
Experimental setup
Datasets: TotalSegmentator TotalSeg++ FLARE22 MSD-Pancreas MSD-Colon
Benchmarks: TotalSeg++ FLARE22 MSD-Pancreas MSD-Colon
Metrics: DSC NSD latency GPU memory GFLOPs
Baselines: SAM MedSAM SAMed MA-SAM SAM-Med2D SAM-Med3D SegVol 3DSAM-adapter nnU-Net UNETR Swin UNETR nnFormer
Builds on: Segment Anything Model TotalSegmentator ImJoy ResT
Closest papers by meaning
- Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.88)
- Unleashing the Potential of SAM for Medical Adaptation via Hierarchical Decoding (CVPR 2024 · similarity 0.88)
- CAT: Coordinating Anatomical-Textual Prompts for Multi-Organ and Tumor Segmentation (NeurIPS 2024 · similarity 0.87)
- VISTA3D: A Unified Segmentation Foundation Model For 3D Medical Imaging (CVPR 2025 · similarity 0.86)
- Point-SAM: Promptable 3D Segmentation Model for Point Clouds (ICLR 2025 · similarity 0.86)
- Better Tokens for Better 3D: Advancing Vision-Language Modeling in 3D Medical Imaging (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.