원클릭으로 Manus에서 모든 스킬 실행

habit-quickstart

스타6

포크2

업데이트2026년 7월 21일 10:36

Entry-point router for the HABIT (Habitat Analysis Biomedical Imaging Toolkit) package. Use this FIRST whenever a user mentions HABIT, habitat analysis, tumor sub-region clustering, intra-tumor heterogeneity, "生境分析", "亚区聚类", "habitat segmentation", or wants a multi-step radiomics workflow. Verifies environment, validates data layout, then hands off to the right specialist skill.

설치

Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.

Manus에서 실행

출처

lichao312214129

lichao312214129/HABIT

GitHub 저장소 열기 Creator 저장소 보기

다운로드

Manus에서 실행

HABIT Quickstart — Workflow Router

This is the entry point for the HABIT package. It does NOT execute analyses on its own — its job is to (1) verify the environment, (2) validate the user's data layout, (3) ask the right clarifying questions, and (4) hand off to the correct specialized skill or recipe.

中文用户提示：本 skill 是 HABIT 工具包的总入口/向导。当用户说"做生境分析"、 "做亚区聚类"、"分析肿瘤异质性"、"我有 MRI 想做 habitat" 等需求时，先读本文件走完路由流程，再去调用对应子 skill 或 recipe。

Step 0 — Environment check (MANDATORY first action)

Before doing anything else, run:

python skills/habit-quickstart/scripts/check_environment.py

This verifies:

Python ≥ 3.8 (≥ 3.10 needed for AutoGluon)
habit CLI is installed and on PATH
All critical packages (SimpleITK, pandas, numpy, sklearn, click, PyYAML) import
Reports any missing optional packages (PyRadiomics, ANTsPy, XGBoost, ...)

If exit code is 1 (failure), tell the user to install HABIT:

conda create -n habit python=3.10
conda activate habit
pip install -e .

If exit code is 2 (warnings only), proceed but warn the user about missing optional packages they'll need later.

Step 1 — Data layout check

If the user mentions a data directory:

python skills/habit-quickstart/scripts/check_data_layout.py <data_dir> --modalities <list>

This catches missing modalities and bad folder layouts BEFORE writing any config. See references/data_layout_spec.md for the canonical layout.

Step 2 — Required Information

Read references/required_questions.md for the per-skill checklist. Do NOT proceed if any required field is unanswered — stop and ask the user. This is per project rule 9 (analyze + propose + get approval first).

Step 3 — Route to the right skill

Use the decision tree in references/workflow_decision_tree.md and this quick table:

User goal	Hand off to
Convert raw DICOM → NIfTI	`habit-preprocess` (dcm2nii mode)
Align / normalize / N4-correct images	`habit-preprocess`
Discover tumor sub-regions	`habit-habitat-analysis`
Extract MSI / ITH / habitat features	`habit-feature-extraction`
Classical PyRadiomics only (no habitat)	`habit-radiomics`
Train / predict / k-fold a classifier	`habit-machine-learning`
Compare ROC / DCA / DeLong of multiple models	`habit-model-comparison`
ICC / test-retest / Dice / merge CSV / DICOM info	`habit-dicom-tools`
End-to-end pipeline	`habit-recipes`
Got an error / something broke	`habit-troubleshoot`

Step 4 — Decision: one_step vs two_step habitat

A common confusion. Default to one_step unless the user is doing a publication-grade multi-cohort study.

one_step — each tumor finds its own optimal cluster number. Faster, simpler, no population-level model. Good for pilots.
two_step — produces a population-level habitat model that generalizes to new patients via --mode predict. Required for predictive modeling studies that need external validation.

Detail: habit-habitat-analysis/references/cluster_validity_guide.md.

Critical project-wide rules (apply to all HABIT work)

All plot/figure outputs use English labels — project-wide rule. Never put Chinese characters on figures. (Code comments / documentation may be Chinese.)
YAML configs: 2-space indent, no tabs, space after colon.
Paths in configs can be relative (resolved against config file location) or absolute.
Always validate after each step — use validate_*.py scripts in each skill's scripts/ folder.
Demo data is available — if the user is new, suggest habit-recipes/references/recipe_demo_walkthrough.md first.

When to stop and ask the user

Do NOT proceed silently if:

User did not specify which images: modalities they have
User did not specify the binary label_col for ML
User wants predict mode but did not provide a trained pipeline path
DICOM-to-NIfTI conversion needs dcm2niix.exe path
User has not picked one_step vs two_step (and you can't safely default)

Resources

Full CLI reference: docs/source/cli_zh.rst
Configuration encyclopedia: docs/source/configuration_zh.rst
Online docs: https://lichao312214129.github.io/HABIT
Demo dataset: see README.md Step 2
YAML path index (config/<module>/ templates): skills/CONFIG_SOURCES.md

이 저장소의 다른 Skills

같은 저장소

habit-troubleshoot

lichao312214129/HABIT

Diagnose and fix HABIT runtime errors — covers preprocess, habitat clustering, feature extraction, and ML/comparison failures. Use when the user pastes a Python traceback, when a `habit ...` CLI command fails, when an output file is missing or corrupt, or when results look biologically wrong (degenerate clusters, all-NaN features, AUC=0.5). Triggers on phrases like "报错", "出错", "不工作", "为什么", "error", "failed", "traceback", "crash", "AUC 太低", "没有生成结果".

2026-07-216

habit-feature-extraction

lichao312214129/HABIT

Extract quantitative features from HABIT habitat maps — traditional radiomics, whole-habitat radiomics, per-habitat radiomics, MSI (Most Significant Intensity), and ITH (Intra-Tumor Heterogeneity) scores. Use when the user has habitat .nrrd maps and wants per-subject feature CSVs ready for ML. Triggers on "提取生境特征", "MSI 特征", "ITH 异质性", "habitat features", "extract radiomics from habitats". Runs `habit extract`.

2026-07-066

habit-habitat-analysis

lichao312214129/HABIT

Generate tumor habitat (sub-region) maps from medical images using voxel-level clustering. Use when the user wants to identify intra-tumor heterogeneity zones, perform supervoxel clustering, generate habitat .nrrd maps, run kinetic DCE habitat, or use voxel-radiomics texture clustering. Triggers on "生境分析", "亚区聚类", "habitat", "supervoxel", "kinetic", "DCE 分期生境", "intra-tumor heterogeneity". Runs `habit get-habitat`.

2026-07-066

habit-machine-learning

lichao312214129/HABIT

Train, predict, or k-fold cross-validate ML classifiers on habitat / radiomics feature CSVs using HABIT. Supports LogisticRegression, RandomForest, XGBoost, SVM, MLP, AutoGluon, and 12+ feature selection methods (LASSO, mRMR, RFECV, ICC, correlation, ANOVA, ...). Use when the user wants to build a prediction model from feature CSVs. Triggers on "训练模型", "建模", "K折交叉验证", "k-fold", "LASSO", "feature selection", "predict mode". Runs `habit model` or `habit cv`.

2026-07-066

habit-model-comparison

lichao312214129/HABIT

Compare multiple trained classification models with publication-quality plots — ROC, DCA, calibration, precision-recall, DeLong's AUC test. Use when the user has prediction CSVs from 2+ models and wants side-by-side comparison. Triggers on "模型比较", "ROC 对比", "DeLong 检验", "决策曲线", "校准曲线", "model comparison", "compare AUC", "DCA". Runs `habit compare`.

2026-07-066

habit-radiomics

lichao312214129/HABIT

Extract traditional PyRadiomics features (firstorder, shape, GLCM, GLRLM, GLSZM, NGTDM, GLDM) from medical images at the whole-tumor ROI level — no habitat segmentation. Use when the user wants classical radiomics features without doing habitat analysis. Triggers on "传统影像组学", "PyRadiomics", "radiomics features", "shape features", "GLCM". Runs `habit radiomics`.

2026-07-066

name	habit-quickstart
description	Entry-point router for the HABIT (Habitat Analysis Biomedical Imaging Toolkit) package. Use this FIRST whenever a user mentions HABIT, habitat analysis, tumor sub-region clustering, intra-tumor heterogeneity, "生境分析", "亚区聚类", "habitat segmentation", or wants a multi-step radiomics workflow. Verifies environment, validates data layout, then hands off to the right specialist skill.

HABIT Quickstart — Workflow Router

中文用户提示：本 skill 是 HABIT 工具包的总入口/向导。当用户说"做生境分析"、 "做亚区聚类"、"分析肿瘤异质性"、"我有 MRI 想做 habitat" 等需求时，先读本文件走完路由流程，再去调用对应子 skill 或 recipe。

Step 0 — Environment check (MANDATORY first action)

Before doing anything else, run:

python skills/habit-quickstart/scripts/check_environment.py

This verifies:

Python ≥ 3.8 (≥ 3.10 needed for AutoGluon)
habit CLI is installed and on PATH
All critical packages (SimpleITK, pandas, numpy, sklearn, click, PyYAML) import
Reports any missing optional packages (PyRadiomics, ANTsPy, XGBoost, ...)

If exit code is 1 (failure), tell the user to install HABIT:

conda create -n habit python=3.10
conda activate habit
pip install -e .

If exit code is 2 (warnings only), proceed but warn the user about missing optional packages they'll need later.

Step 1 — Data layout check

If the user mentions a data directory:

python skills/habit-quickstart/scripts/check_data_layout.py <data_dir> --modalities <list>

This catches missing modalities and bad folder layouts BEFORE writing any config. See references/data_layout_spec.md for the canonical layout.

Step 2 — Required Information

Step 3 — Route to the right skill

Use the decision tree in references/workflow_decision_tree.md and this quick table:

User goal	Hand off to
Convert raw DICOM → NIfTI	`habit-preprocess` (dcm2nii mode)
Align / normalize / N4-correct images	`habit-preprocess`
Discover tumor sub-regions	`habit-habitat-analysis`
Extract MSI / ITH / habitat features	`habit-feature-extraction`
Classical PyRadiomics only (no habitat)	`habit-radiomics`
Train / predict / k-fold a classifier	`habit-machine-learning`
Compare ROC / DCA / DeLong of multiple models	`habit-model-comparison`
ICC / test-retest / Dice / merge CSV / DICOM info	`habit-dicom-tools`
End-to-end pipeline	`habit-recipes`
Got an error / something broke	`habit-troubleshoot`

Step 4 — Decision: one_step vs two_step habitat

A common confusion. Default to one_step unless the user is doing a publication-grade multi-cohort study.

one_step — each tumor finds its own optimal cluster number. Faster, simpler, no population-level model. Good for pilots.
two_step — produces a population-level habitat model that generalizes to new patients via --mode predict. Required for predictive modeling studies that need external validation.

Detail: habit-habitat-analysis/references/cluster_validity_guide.md.

Critical project-wide rules (apply to all HABIT work)

All plot/figure outputs use English labels — project-wide rule. Never put Chinese characters on figures. (Code comments / documentation may be Chinese.)
YAML configs: 2-space indent, no tabs, space after colon.
Paths in configs can be relative (resolved against config file location) or absolute.
Always validate after each step — use validate_*.py scripts in each skill's scripts/ folder.
Demo data is available — if the user is new, suggest habit-recipes/references/recipe_demo_walkthrough.md first.

When to stop and ask the user

Do NOT proceed silently if:

User did not specify which images: modalities they have
User did not specify the binary label_col for ML
User wants predict mode but did not provide a trained pipeline path
DICOM-to-NIfTI conversion needs dcm2niix.exe path
User has not picked one_step vs two_step (and you can't safely default)

Resources

Full CLI reference: docs/source/cli_zh.rst
Configuration encyclopedia: docs/source/configuration_zh.rst
Online docs: https://lichao312214129.github.io/HABIT
Demo dataset: see README.md Step 2
YAML path index (config/<module>/ templates): skills/CONFIG_SOURCES.md