segmentation
Guides Attention-Gated Networks medical image segmentation workflows, including 3D U-Net training, validation, NIfTI layouts, and attention or feature-map export.
- License 许可
- MIT
- 最近更新
- 2026-10-08
- Stars
- ★ 11
- Forks
- 3
01 前言 Front Matter
- name
- segmentation
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Guides Attention-Gated Networks medical image segmentation workflows, including 3D U-Net training, validation, NIfTI layouts, and attention or feature-map export.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/api-reference.md6.1 KB
- 参考文档references/data-layout.md3.2 KB
- 参考文档references/troubleshooting.md4.6 KB
- 参考文档references/workflows.md4.1 KB
03 SKILL.md 正文
Segmentation
Use this sub-skill for medical image segmentation tasks in Attention-Gated Networks. It covers 2D/3D U-Net variants, CT deep-supervision networks, attention-gated U-Nets, non-local blocks, NIfTI datasets, segmentation training/validation scripts, and feature or attention map exports.
Read first
- Read api-reference.md when you need exact model names, wrapper methods, dataset class signatures, metric functions, or hookable layer names.
- Read data-layout.md before arranging NIfTI data or debugging split, image/label, and transform problems.
- Read workflows.md for training, validation, checkpoint use, and safe attention/feature map export.
- Read troubleshooting.md for CUDA, memory, SimpleITK, torchsample, NIfTI, shape, metric, or layer-hook failures.
- Run run_segmentation.py for skill-owned training commands that replace the source repo's root training entry point.
- Run validate_and_export_maps.py to replace hard-coded validation and feature-map scripts with a parameterized, skill-owned helper.
- For fast environment validation, run the root helper
../../scripts/check_env.py --repo-root /path/to/Attention-Gated-Networks --mode segmentation.
When to use this sub-skill
Use this route when the user asks to:
- train a segmentation network from the repository configs;
- choose between baseline U-Net, non-local U-Net, CT deep supervision, and multi-attention U-Net variants;
- validate a segmentation checkpoint and compute dice, IoU, precision/recall, mean distance, or Hausdorff-style distances;
- arrange 3D NIfTI data for
acdc_sax,ukbb_sax,rvsc_sax, ortest_sax; - export NIfTI predictions, attention volumes, feature maps, or slice previews;
- debug 3D shape, memory, CUDA, SimpleITK, torchsample, or layer-hook issues.
Route ultrasound scan-plane classification and 2D attention overlays to
../classification/SKILL.md instead.
Quick workflow
-
Install the repository with CUDA-enabled PyTorch, torchsample, NIfTI, and visualization dependencies. Then run:
python ../../scripts/check_env.py --repo-root /path/to/Attention-Gated-Networks --mode segmentation -
Arrange NIfTI files under split directories with paired
image/andlabel/subfolders. See data-layout.md. -
Copy a CT config and replace
data_path.acdc_saxwith the current dataset root. The stock configs useoutput_nc=4,tensor_dim='3D', andgpu_ids=[0].
Relative config paths are resolved from the explicit --repo-root; relative
data_path values are resolved from the config file's parent. NIfTI data and
model weights/checkpoints must exist externally. The bundled runner and
check_env.py --config ... fail fast on missing or private /vol/... paths;
they never fabricate data.
4. Train with the bundled replacement:
python scripts/run_segmentation.py --repo-root /path/to/Attention-Gated-Networks --config path/to/config.json
-
Validate a checkpoint or export maps with the bundled helper:
python scripts/validate_and_export_maps.py --repo-root /path/to/Attention-Gated-Networks --config path/to/config.json --checkpoint path/to/checkpoint.pth --output-dir /tmp/ag-net-validation --mode validate -
For safe smoke validation or map export without private paths, use:
python scripts/validate_and_export_maps.py \ --repo-root /path/to/Attention-Gated-Networks \ --config configs/config_unet_ct_multi_att_dsv.json \ --output-dir /tmp/ag-net-segmentation \ --mode both
Key decisions
- CUDA is required for the unmodified source wrappers. A CPU-only install does not verify the selected segmentation workflows.
unet_ct_dsvis the simpler CT deep-supervision route; useunet_ct_multi_att_dsvwhen attention maps are a first-class goal.- Real CT configs use large volumes (
160 x 160 x 96patches). The generated helper's16 x 16 x 16synthetic volume is only a wiring smoke. - The source validation and map-export scripts hard-code private paths. Prefer the bundled helper or explicitly parameterize a project-local copy.
- The source CRF post-processing script is reference-only, not bundled as a safe helper.
Expected outputs
Training writes checkpoints and logs below the configured checkpoint directory
and experiment name. Validation writes NIfTI predictions and CSV metrics. The
bundled helper writes synthetic or fixture NIfTI outputs, metrics.json,
feature/attention-map NIfTI files, and middle-slice preview PNGs to the selected
output directory.
04 同仓库更多 Skills
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
python-api
Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP/mAP semantics, and source-style import diagnostics.
file-evaluation
Folder-based PASCAL VOC AP/mAP evaluation from ground-truth and detection text files.
object-detection-metrics
Guides legacy Object-Detection-Metrics PASCAL VOC AP/mAP evaluation from detection text folders and direct Python API classes.
nx-workspace-patterns
Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.