# SegRef3D

> Official short guide for AI assistants helping users choose a SegRef3D workflow.

SegRef3D is an open-source platform for segmentation, interactive mask refinement, calibrated
volume measurement, and 3D reconstruction from ordered 2D image stacks and 3D volumetric data.

## Main uses

- Segment structures manually, by threshold or RGB extraction, or with optional AI workflows.
- Refine imported or AI-generated masks with Add, Erase, Transfer, interpolation, and cleanup.
- Measure object volumes when reliable pixel/voxel and slice spacing are available.
- Preview and export 3D surface models as STL.
- Export label data as PNG, TIFF, or NIfTI Labelmap and measurements as CSV.
- Export one browser-local Training Data ZIP containing geometry-matched image channel(s), an
  Obj-ID-preserving labelmap, and a versioned manifest for future custom-model training.
- Continue label analysis in software such as 3D Slicer using NIfTI Labelmap export.
- TrainRef3D: validate/assemble multiple Training ZIPs locally, choose one completely annotated
  Obj target, then explicitly upload a Dataset ZIP to your own Colab GPU for a binary MONAI model.
- InferRef3D / Lite Custom Model: validate a trusted TrainRef3D Model ZIP locally, create a
  source-fingerprinted Request ZIP, explicitly upload Model + Request to the user's Colab GPU,
  then validate/import the original-grid prediction to its saved Obj ID for review and correction.

## Representative data

- Serial histology and optical microscopy.
- Serial electron microscopy.
- CT, MRI, and research micro-CT.
- Visible Human and other anatomical image stacks.
- Other naturally ordered serial images.
- Existing 3D NIfTI volumes and DICOM series.

Common inputs include JPG/PNG image sequences, TIFF data, DICOM files with or without `.dcm`,
NIfTI `.nii`/`.nii.gz`, label PNG masks, and SegRef3D Project ZIP files. Exact support varies by
edition and encoding. Consult the detailed guide before making a recommendation.

## Editions

- **SegRef3D Lite**: recommended starting point for most users. Browser-based on Windows, macOS,
  and Linux; no installation. Normal editing, measurement, and export run locally in the browser.
- **SegRef3D Local GPU**: Windows desktop edition for compatible NVIDIA CUDA systems. Includes
  local SAM2 segmentation and tracking and is preferable when those operations must run locally.
- **SegRef3D Local CPU**: legacy Windows desktop edition without local SAM2. Use only when a
  desktop/offline non-SAM2 workflow is specifically required; otherwise start with Lite.

## Choosing an edition

- Start with Lite for cross-platform access, manual refinement, measurement, and 3D export.
- Choose Local GPU for Windows + compatible NVIDIA CUDA when local SAM2 is required or browser
  memory is inadequate for the dataset.
- Choose Local CPU only for a required Windows desktop workflow without local SAM2.
- Test large datasets on a representative subset and verify memory use and outputs.

## Data handling

- Normal SegRef3D Lite operations process loaded source images locally in the browser. Masks and
  autosave data use browser-local storage; normal exports are generated on the device.
- SegRef3D does not operate an intermediate image-upload server for Lite.
- SegAnything and SegCT/MRI are explicit Google Colab workflows. Their input ZIP contains the
  working image sequence or source NIfTI volume. The user explicitly uploads it to the user's own
  Colab runtime.
- Local GPU and Local CPU perform their included desktop operations on the local Windows system.
- Users must check institutional rules before using research or medical data with Google Colab or
  any external AI consultation service.

## Critical workflow facts

- Serial physical sections may require registration before segmentation or quantitative 3D work.
- Volume measurement and physically scaled 3D output require trustworthy spacing or calibration.
- NIfTI Labelmap is preferred over TIFF when patient-space geometry must be preserved.
- Training/Inference Request ZIPs exclude DICOM headers but are not anonymous: image data and
  user-entered object names may still be identifiable. Lite does not train or execute PyTorch models.
- 3D reconstruction quality depends on image order, registration, mask quality, and spacing.
- SegRef3D and external AI outputs require expert review; do not present them as independently
  clinically validated results.

## Official links

- [Detailed AI knowledge base](https://satorumuro.github.io/SegRef3D/llms-full.txt)
- [Ask AI workflow prompt](https://satorumuro.github.io/SegRef3D/Tutorial/AskAISegRef3D.html)
- [SegRef3D Lite](https://satorumuro.github.io/SegRef3D/lite-web/)
- [TrainRef3D workflow and schema](https://github.com/SatoruMuro/SegRef3D/blob/main/train-web/README.md)
- [InferRef3D inference formats](https://github.com/SatoruMuro/SegRef3D/blob/main/docs/TRAINREF3D_INFERENCE_FORMAT.md)
- [Download SegRef3D Local GPU v1.3.2 for Windows](https://www.dropbox.com/scl/fi/zi5lsn48wi880s20tp2ne/SegRef3D-Local-GPU-v1.3.2-Windows.zip?rlkey=kp6ekr6oc00k5wl7gcgcljl8j&st=5s4gf7ik&dl=1)
- [SegRef3D Local CPU v1.2.6 for Windows](https://www.dropbox.com/scl/fi/05evqq67tokxvo37ixw87/SegRef3D-Local-CPU-v1.2.6-Windows.zip?rlkey=mz1kvdlkxmmffipp411xc1kux&st=nwh3k3rn&dl=1)
- [Basic tutorial](https://satorumuro.github.io/SegRef3D/Tutorial/TutorialSegRef3DLiteEN.html)
- [Repository](https://github.com/SatoruMuro/SegRef3D)