Back to Details

imaging-study-design

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Imaging Study Design and Biomarker Selection

医学影像研究与生物标志物选择

Overview

概述

This skill encodes methodology for designing medical imaging studies and selecting imaging biomarkers. It guides decisions about preprocessing, registration targets, de-identification, biomarker choice, radiomics stability, and multi-site harmonization. It is a reasoning skill — it does not generate pipeline code. For concrete tooling (FreeSurfer, FSL, ANTs, PyRadiomics, nnU-Net, MONAI), pair this skill with a pipeline skill after the study design is settled.
The central principle: preprocessing and biomarker choice must preserve the signal that answers the clinical question. A pipeline that is standard for one question (e.g., MNI registration for group comparison) can destroy the signal for another (e.g., individual atrophy in longitudinal studies).
本 skill 系统化整理了医学影像研究设计和影像生物标志物选择的方法。它指导如何进行预处理、选择配准目标、去标识化、选择生物标志物、评估 radiomics 稳定性以及开展多中心 harmonization。它是一种推理 skill，不生成 pipeline 代码。对于具体工具（FreeSurfer、FSL、ANTs、PyRadiomics、nnU-Net、MONAI），应在研究设计确定后，将本 skill 与相应的 pipeline skill 配合使用。
核心原则：预处理和生物标志物选择必须保留能够回答临床问题的信号。一个适合某类问题的标准流程（例如，为组间比较进行 MNI 配准）可能会破坏另一类问题中的信号（例如，纵向研究中的个体萎缩）。

Usage

使用方法

Invoke this skill when planning an imaging study or critiquing an existing imaging pipeline. Typical prompts:
  • "I want to measure hippocampal atrophy over 2 years in Alzheimer's patients — how should I preprocess?"
  • "Design a multi-site radiomics study for NSCLC prognosis."
  • "We're collecting brain MRI across three scanners. What harmonization do I need?"
  • "Is it safe to share these DICOMs after standard de-identification?"
  • "Which imaging biomarker should I use for presurgical mapping in epilepsy?"
Use the Decision Framework at the end of Core Concepts to structure any imaging study question.
在规划影像研究或评审现有影像 pipeline 时调用本 skill。典型提示词包括：
  • “我想测量阿尔茨海默病患者 2 年内的海马萎缩——应如何预处理？”
  • “为 NSCLC 预后设计一项多中心 radiomics 研究。”
  • “我们正在使用三台扫描仪采集脑 MRI，需要采用哪种 harmonization？”
  • “完成标准去标识化后，安全共享这些 DICOM 吗？”
  • “用于癫痫术前定位，应选择哪种影像生物标志物？”
使用“核心概念”末尾的决策框架组织任何影像研究问题。

Response Format

响应格式

  • Lead with the direct recommendation or classification (≤3 sentences)
  • Structure as: recommendation → justification (citing specific criteria/thresholds) → caveats
  • Use tables for comparisons; bullet points for criteria lists
  • Omit background the user already knows — they asked the question
  • Target: 200-400 words unless the user requests exhaustive detail The decision trees and frameworks in this skill are for internal reasoning only. Apply them to reach your conclusion, but do not reproduce them in your response. Present only the final recommendation with supporting evidence.
  • 首先给出直接建议或分类结论（不超过 3 句）
  • 按以下结构组织：建议 → 理由（引用具体标准或阈值）→ 注意事项
  • 比较使用表格；标准清单使用项目符号
  • 省略用户已经了解的背景——他们是在提问
  • 除非用户要求详尽说明，目标长度为 200–400 字
本 skill 中的决策树和框架仅用于内部推理。应应用这些框架得出结论，但不要在回答中复现。只呈现最终建议及其支持证据。

Core Concepts

核心概念

1. Preprocessing strategy is downstream-goal-dependent

1. 预处理策略取决于下游目标

There is no universal "preprocess MRI" recipe. The pipeline must match the analysis target.
Downstream goalPreprocessing prioritiesWhat to avoid
Group comparison (cross-sectional)Register to common template (MNI/SPM), smooth, modulate if VBMNative-space-only analysis
Longitudinal atrophy / changeRigid register follow-ups to subject baseline, then (optionally) baseline to templateRegistering each timepoint independently to MNI
Lesion segmentationBias field correction, skull strip, intensity normalization; registration often optionalAggressive smoothing that blurs lesion edges
Radiomics / textureStandardize acquisition first (voxel size, reconstruction, binning); resample to isotropic; discretize intensity deterministicallyMixing protocols without harmonization; default-bin radiomics on heterogeneous data
Functional MRISlice-timing (if long TR), motion correction, distortion correction, coregister to anatomical, normalize if group, bandpass filter as appropriateNormalizing before motion scrubbing; smoothing before ICA denoising
Diffusion MRIDenoise, Gibbs unringing, eddy + motion correction, bias correction, then tensor/ODF fittingReordering these steps; skipping distortion correction when reverse-PE is available
不存在通用的“MRI 预处理”方案。pipeline 必须与分析目标相匹配。
下游目标预处理重点应避免
组间比较（横断面）配准至通用模板（MNI/SPM），进行平滑；若使用 VBM 则进行调制仅在原生空间进行分析
纵向萎缩或变化将随访扫描刚性配准至受试者基线，随后可选地将基线配准至模板将每个时间点独立配准到 MNI
病灶分割偏置场校正、去颅骨、强度归一化；配准通常并非必需过度平滑导致病灶边缘模糊
Radiomics 或纹理分析首先统一采集参数（体素大小、重建方式、强度分箱）；重采样至各向同性；以确定性方式离散化强度混合 protocol 却不进行 harmonization；在异质数据上使用默认分箱进行 radiomics
功能 MRI时间切片校正（若 TR 较长）、运动校正、畸变校正、与解剖像 coregister，如进行组分析则归一化，必要时进行带通滤波在 motion scrubbing 前进行归一化；在 ICA 去噪前进行平滑
Diffusion MRI去噪、Gibbs ringing 校正、eddy + 运动校正、偏置校正，然后进行 tensor/ODF 拟合调整这些步骤的顺序；在有 reverse-PE 数据时跳过畸变校正

2. Registration target decisions

2. 配准目标决策

The registration target encodes an assumption about where the signal lives.
  • MNI / population template — for cross-sectional group statistics, voxelwise analyses (VBM, TBSS), atlas-based ROI extraction. Accepts that individual anatomy is partly warped away.
  • Subject baseline (within-subject template) — for longitudinal analyses of change (atrophy, tumor growth, lesion evolution). Preserves each subject's anatomy and measures change in a consistent individual frame.
  • Native space — for lesion analysis, surgical planning, radiomics on tumors, any case where registration would deform the structure of interest or introduce interpolation artifacts in the ROI.
  • Subject anatomical (T1) — for within-subject cross-modal alignment (fMRI, DWI, PET to T1) before any group normalization.
Heuristic: ask "if I warp this image, do I still measure the thing I care about?" If the answer is no (volume change, lesion shape, tumor texture), keep the measurement in native or within-subject space.
配准目标体现了对信号所在位置的假设。
  • MNI 或群体模板——适用于横断面组间统计、体素级分析（VBM、TBSS）和基于图谱的 ROI 提取。需要接受个体解剖结构被部分形变消除这一代价。
  • 受试者基线（受试者内模板）——适用于变化情况的纵向分析（萎缩、肿瘤生长、病灶演变）。保留每位受试者的解剖结构，并在一致的个体坐标系中测量变化。
  • 原生空间——适用于病灶分析、手术规划、肿瘤 radiomics，以及配准会扭曲目标结构或在 ROI 中引入插值伪影的任何情况。
  • 受试者解剖像（T1）——用于任何组水平归一化之前的受试者内跨模态对齐（fMRI、DWI、PET 到 T1）。
启发式判断标准：“如果我对这张图像进行形变配准后，还能测量所关心的对象吗？”如果答案是不能（如体积变化、病灶形状或肿瘤纹理），则应在原生空间或受试者内空间中进行测量。

3. DICOM de-identification risk surface

3. DICOM 去标识化风险面

Standard tag-based de-identification (DICOM PS 3.15 Basic Application Level Confidentiality Profile) is necessary but not sufficient. Residual PHI risks:
  • Burned-in annotations in pixel data — ultrasound frames, screen captures, secondary captures, dose reports, 3D reformats. Patient name/MRN/DOB is often rendered into the pixels. Tag-level scrubbing does not remove this. Requires visual QC or OCR-based pixel redaction.
  • Private tags by manufacturer — Siemens CSA headers (
    0029,xx10
    / 
    0029,xx20
    ) may contain protocol and sometimes patient info; GE private blocks (e.g., 
    0043,xx
    ) and Philips MR private tags can hold acquisition metadata that includes identifiers or free-text. Default DICOM profiles may retain or only partially scrub these. Decide per-vendor: strip all private tags unless a specific block is needed, and validate.
  • Structured reports (SR) and encapsulated PDFs — narrative findings often contain patient name, referring physician, dates, prior history. Must be treated as free-text and redacted or dropped.
  • UIDs encoding dates or MRNs — some sites generate Study/Series/SOP Instance UIDs from timestamps or record numbers. Rewrite UIDs with a consistent hash (preserving referential integrity within the release) rather than leaving originals.
  • Facial reconstructability from head MRI/CT — high-resolution T1/CT allow face re-rendering. Apply defacing (e.g., remove or deform voxels around the face) for public release; consider skull-stripping for research-only shares.
  • Acquisition dates and times — even after name removal, exact dates can re-identify when combined with auxiliary data. Date-shift per subject with a preserved offset if relative timing matters.
A defensible de-identification plan specifies: which profile is applied, how private tags are handled, whether pixel data is inspected, how SR/PDF objects are handled, UID remapping strategy, defacing policy, and a QC sample checked by a human.
标准的基于标签的去标识化（DICOM PS 3.15 Basic Application Level Confidentiality Profile）是必要条件，但并不充分。残余 PHI 风险包括：
  • 像素数据中烧录的标注——超声帧、屏幕截图、二次采集图像、剂量报告和 3D 重建图像。患者姓名、MRN 或 DOB 通常已渲染到像素中。标签级清理无法移除这些信息，必须进行目视 QC 或基于 OCR 的像素脱敏。
  • 厂商私有标签——Siemens CSA 头（
    0029,xx10
    / 
    0029,xx20
    ）可能包含 protocol 信息，有时还包含患者信息；GE 私有数据块（如 
    0043,xx
    ）和 Philips MR 私有标签可能保存包含标识符或自由文本的采集元数据。默认 DICOM 配置可能保留或仅部分清理这些标签。应按厂商分别制定策略：除非确需某个特定数据块，否则删除所有私有标签，并进行验证。
  • 结构化报告（SR）和封装 PDF——叙述性检查结果通常包含患者姓名、转诊医生、日期和既往病史。必须将其视为自由文本进行脱敏或删除。
  • 编码日期或 MRN 的 UID——某些中心会使用时间戳或病历编号生成 Study、Series 或 SOP Instance UID。应使用一致的哈希值重写 UID，同时在发布包中保持引用完整性，而不是保留原始值。
  • 通过头部 MRI/CT 重建面部——高分辨率 T1 或 CT 可用于重建面部。公开发布时应进行 defacing（例如移除或变形面部周围的体素）；仅限研究用途的共享可考虑去颅骨。
  • 采集日期和时间——即使姓名已删除，精确日期与辅助数据结合后仍可能实现再识别。若相对时间关系很重要，可按受试者平移日期，同时保留时间差。
一项可辩护的去标识化方案应明确：采用哪个配置文件、如何处理私有标签、是否检查像素数据、如何处理 SR/PDF 对象、UID 重映射策略、defacing 策略，以及由人工检查的 QC 样本。

4. Imaging biomarker selection by indication

4. 按适应证选择影像生物标志物

Match the biomarker to the biology being probed.
IndicationModality / contrastBiomarkerRationale
Alzheimer's disease, MCIStructural T1 MRIHippocampal volume, cortical thickness, ventricular volumeMeasures neurodegeneration directly
Multiple sclerosisT2/FLAIR + T1Lesion load, lesion count, brain parenchymal fraction, cord areaDisease burden and atrophy
White matter disease, DAI, schizophreniaDiffusion MRIFA, MD, RD, AD (tract-based or voxelwise)Microstructural integrity
Acute strokeDWI + PWIDWI lesion volume, PWI-DWI mismatch, CBFCore vs penumbra triage
Brain tumor (glioma)DSC/DCE perfusion, MRS, diffusionrCBV, Ktrans, ADC, NAA/Cho ratioGrade, treatment response, pseudoprogression vs recurrence
Presurgical planning (epilepsy, tumor)Task and resting-state fMRI, DTI tractographyLanguage/motor activation maps, tract-to-lesion distanceFunctional localization
Neuronal integrity (various)1H-MRSNAA/Cr, Cho/Cr, mI/CrMetabolic markers of neuronal loss, membrane turnover
Oncology outside brainCT, PET/CT, mpMRIVolumetry, SUVmax/SUVpeak, radiomic signatures, ADCStaging, response (RECIST, PERCIST)
CardiacCine, LGE, T1/T2 mappingEF, strain, LGE burden, native T1, ECVFunction, fibrosis, edema
Selection criteria: (1) biological plausibility — is this biomarker mechanistically linked to the question? (2) measurability — does the available scanner/protocol yield adequate SNR and reliability? (3) validity — has it been validated for this indication and this population? (4) practicality — can it be obtained within the study's scan time and cost?
应使生物标志物与所检测的生物学机制相匹配。
适应证模态或对比生物标志物依据
阿尔茨海默病、MCI结构 T1 MRI海马体积、皮层厚度、脑室体积直接测量神经退行性变
多发性硬化T2/FLAIR + T1病灶负荷、病灶数量、脑实质比例、脊髓面积疾病负荷与萎缩
白质疾病、DAI、精神分裂症Diffusion MRIFA、MD、RD、AD（基于纤维束或体素级）微观结构完整性
急性卒中DWI + PWIDWI 病灶体积、PWI-DWI 不匹配、CBF核心与半暗带分诊
脑肿瘤（胶质瘤）DSC/DCE 灌注、MRS、DiffusionrCBV、Ktrans、ADC、NAA/Cho 比值分级、治疗反应、假性进展与复发鉴别
术前规划（癫痫、肿瘤）任务态和静息态 fMRI、DTI 纤维束追踪语言或运动激活图、病灶至纤维束距离功能定位
神经元完整性（多种疾病）1H-MRSNAA/Cr、Cho/Cr、mI/Cr神经元丢失和膜周转的代谢标志物
脑外肿瘤CT、PET/CT、mpMRI体积测量、SUVmax/SUVpeak、radiomic signature、ADC分期、疗效评价（RECIST、PERCIST）
心脏Cine、LGE、T1/T2 mappingEF、应变、LGE 负荷、native T1、ECV功能、纤维化、水肿
选择标准：（1）生物学合理性——该生物标志物是否与研究问题存在机制层面的联系？（2）可测性——现有扫描仪或 protocol 能否提供足够的 SNR 和可靠性？（3）效度——该生物标志物是否已针对此适应证和此类人群得到验证？（4）可行性——能否在研究规定的扫描时间和成本内获取？

5. Radiomics stability and sample-size discipline

5. Radiomics 稳定性与样本量约束

Radiomics features are notoriously unstable. A signature built on unstable features will not generalize.
  • Test-retest reliability — require ICC ≥ 0.75 (commonly ICC ≥ 0.9 for high confidence) on repeat scans for any feature entering the model. Use a test-retest cohort or a publicly available one (e.g., RIDER Lung CT) aligned to the modality.
  • IBSI compliance — use an IBSI-compliant extractor (PyRadiomics with IBSI settings, CERR, LIFEx) and report IBSI feature definitions and hash. Non-compliant extractors produce features that cannot be reproduced across sites.
  • Segmentation robustness — evaluate inter-rater and intra-rater ICC for feature values under realistic contour variability, not just ideal contours.
  • Acquisition standardization first — fixing voxel size, reconstruction kernel, and intensity discretization (fixed bin width vs fixed bin count — choose one and justify) eliminates more variance than any post-hoc harmonization.
  • Harmonization — apply ComBat (or its variants: parametric/non-parametric, with or without covariate preservation) to remove scanner/site effects after acquisition standardization, not as a substitute for it.
  • Feature reduction before modeling — cluster redundant features (|ρ| > 0.8), drop unstable features, then apply a supervised reducer (LASSO, mRMR) inside cross-validation. Cap the final model at roughly 1 feature per 10 events (EPV ≥ 10) to avoid overfitting. For low-event studies, report this constraint explicitly; it often means choosing 3–5 features, not 20.
  • Report a TRIPOD- or CLEAR-style pipeline description so the study is reproducible.
Radiomics 特征素以不稳定著称。由不稳定特征构建的模型无法泛化。
  • Test-retest reliability——任何进入模型的特征，在重复扫描上均应达到 ICC ≥ 0.75（高置信度通常要求 ICC ≥ 0.9）。应使用 test-retest 队列或与模态匹配的公开队列（如 RIDER Lung CT）。
  • IBSI 兼容性——使用符合 IBSI 的特征提取器（采用 IBSI 设置的 PyRadiomics、CERR、LIFEx），并报告 IBSI 特征定义和哈希值。不兼容的提取器产生的特征无法在不同中心之间复现。
  • 分割稳健性——在符合实际的轮廓变异下评估特征值的评分者间和评分者内 ICC，而不能只评估理想轮廓。
  • 首先统一采集参数——固定体素大小、重建核和强度离散化方式（固定 bin width 还是固定 bin count——选择其一并说明理由），比任何事后 harmonization 都能消除更多变异。
  • Harmonization——在完成采集标准化后，使用 ComBat 或其变体（参数化或非参数化、是否保留协变量）消除扫描仪或中心效应，而不能以此替代采集标准化。
  • 建模前进行特征降维——聚类冗余特征（
    |ρ| > 0.8
    ），删除不稳定特征，然后在交叉验证内部使用监督式降维器（LASSO、mRMR）。最终模型的特征数大致限制为每 10 个事件 1 个特征（EPV ≥ 10），以避免过拟合。对于低事件数研究，应明确报告这一限制；这通常意味着选择 3–5 个特征，而不是 20 个。
  • 报告 TRIPOD 或 CLEAR 风格的 pipeline 描述，以保证研究可复现。

6. Multi-site study considerations

6. 多中心研究注意事项

Multi-site studies trade statistical power for between-site heterogeneity. Plan for the heterogeneity.
  • Protocol standardization — adopt ADNI-style or equivalent harmonized protocols (sequence parameters, resolution, coverage, reconstruction). Agree on inclusion criteria for scanner models and software versions.
  • Scanner and coil effects — fixed by protocol where possible, residual effects modeled with ComBat or linear mixed models (site as random effect). A traveling phantom (one subject or a physical phantom scanned at all sites) quantifies residual bias.
  • QC metrics, scored per scan — SNR, CNR, motion (e.g., FD for fMRI, Euler number for FreeSurfer T1 QC), ghosting, coverage, geometric distortion. Define pre-specified exclusion thresholds before the analysis starts, not after.
  • Central vs local processing — central processing removes local pipeline variance; local processing with a harmonized Docker/Singularity container is acceptable if the container is version-pinned.
  • Reader variability — for any human-in-the-loop step (segmentation, lesion counting, BI-RADS), use ≥ 2 readers on a sample, report Cohen's κ or ICC, and adjudicate disagreements with a predefined rule.
  • Analysis plan — pre-register the statistical model. Decide in advance whether site is a fixed effect (few sites, stable) or random effect (many sites, generalization target), and whether ComBat is applied at feature level or image level.
多中心研究以中心间异质性为代价换取统计功效。必须针对这种异质性进行规划。
  • Protocol 标准化——采用 ADNI 风格或等效的协调 protocol（序列参数、分辨率、覆盖范围、重建方式）。就扫描仪型号和软件版本的纳入标准达成一致。
  • 扫描仪与线圈效应——尽可能通过 protocol 固定这些效应，残余效应则使用 ComBat 或线性混合模型建模（将 site 作为随机效应）。使用 traveling phantom（同一名受试者或在各中心扫描的物理体模）量化残余偏差。
  • 逐次扫描评分的 QC 指标——SNR、CNR、运动程度（如 fMRI 的 FD、FreeSurfer T1 QC 的 Euler number）、鬼影、覆盖范围和几何畸变。应在分析开始前而非分析后预先规定排除阈值。
  • 集中处理与本地处理——集中处理可消除本地 pipeline 差异；如采用统一且版本固定的 Docker/Singularity 容器，本地处理也可以接受。
  • 阅片者差异——对于任何人工参与环节（分割、病灶计数、BI-RADS），应在样本中安排至少 2 名阅片者，报告 Cohen's κ 或 ICC，并按照预先定义的规则裁定分歧。
  • 分析计划——预注册统计模型。预先决定将 site 视为固定效应（中心数量少且稳定）还是随机效应（中心数量多且目标是泛化），并决定 ComBat 应用于特征层面还是图像层面。

7. Decision framework — apply to every imaging study

7. 决策框架——应用于每项影像研究

Before writing a single line of pipeline code, answer four questions:
  1. What is the clinical question? — Screening, diagnosis, staging, prognosis, treatment response, mechanism? The question determines the endpoint and the acceptable error profile.
  2. What modality and contrast answers it? — Match biology to physics (see §4). If multiple modalities answer it, pick the one with the best cost/benefit and existing validation for the indication.
  3. What preprocessing preserves the signal of interest? — Work backward from the measurement. If the measurement is volume change, preserve individual anatomy. If it is group-level activation, register to a template. If it is texture, standardize acquisition and avoid interpolation across the ROI.
  4. What confounds must be controlled? — Age, sex, scanner, site, coil, software version, motion, time of day, medication, scan quality. List them explicitly, decide which are in the model, which are excluded by design, and which are acknowledged as limitations.
If any of the four is unclear or contested, resolve it before building the pipeline. Most imaging study failures are failures of step 1 or 3, not failures of the tool.
在编写任何 pipeline 代码之前，应回答四个问题：
  1. 临床问题是什么？——筛查、诊断、分期、预后、治疗反应还是机制？问题决定研究终点和可接受的误差范围。
  2. 哪种模态和对比能够回答问题？——使生物学机制与物理成像原理相匹配（见 §4）。如果多种模态都能回答问题，应选择成本效益比最佳且已针对该适应证得到验证的模态。
  3. 哪种预处理能够保留目标信号？——从测量目标反向推导。如果测量的是体积变化，应保留个体解剖结构；如果测量的是组水平激活，应配准到模板；如果测量的是纹理，应统一采集参数并避免在 ROI 内进行插值。
  4. 必须控制哪些混杂因素？——年龄、性别、扫描仪、中心、线圈、软件版本、运动、时段、用药和扫描质量。应明确列出这些因素，并分别决定哪些纳入模型、哪些通过设计排除、哪些作为局限说明。
如果四个问题中的任何一个尚不明确或存在争议，应在构建 pipeline 前予以解决。大多数影像研究失败源于第 1 步或第 3 步，而不是工具本身。

When NOT to Use This Skill

不应使用本 skill 的情况

  • Selecting imaging protocols for individual patient diagnosis (clinical radiology)
  • When regulatory submission requires qualified imaging CRO oversight
  • Real-time image interpretation or radiological reporting
  • 为单个患者诊断选择影像 protocol（临床放射学）
  • 监管申报要求具备资质的影像 CRO 监督
  • 实时图像解读或放射学报告

When to Escalate to a Human Expert

需要升级至人工专家时

  • Multi-site harmonization decisions affecting primary endpoints
  • When imaging biomarker will serve as surrogate endpoint in a trial
  • Scanner-specific protocol optimization requiring vendor expertise
  • 影响主要终点的多中心 harmonization 决策
  • 影像生物标志物将作为试验替代终点时
  • 需要厂商专业知识的扫描仪特定 protocol 优化

Common Mistakes

常见错误

  • Wrong: Registering each longitudinal timepoint independently to MNI Right: Build a within-subject template or rigidly register follow-ups to baseline first; only then normalize to MNI if needed Why: Subject-specific atrophy is absorbed into the warp field and the longitudinal change signal is lost
  • Wrong: Relying solely on tag-based DICOM de-identification Right: Include pixel inspection, private-tag policy, SR/PDF handling, UID remapping, and defacing in the de-identification plan Why: Burned-in PHI in pixels, private tags, SR/PDF narratives, and date-encoded UIDs all survive a naive tag-level scrub
  • Wrong: Building a radiomics model using features with test-retest ICC < 0.75 Right: Filter features by ICC ≥ 0.75 (preferably ≥ 0.9) and by segmentation robustness before any supervised selection Why: Unstable features poison reproducibility — signatures built on them will not generalize to external cohorts
  • Wrong: Ignoring scanner/site effects in multi-site imaging studies Right: Standardize protocols, apply ComBat or mixed-effects modeling, include site as a covariate, and pre-register the analysis plan Why: Site confounding can fully explain apparent biological effects, producing false-positive findings
  • Wrong: Fitting dozens of radiomics features to a dataset with few outcome events Right: Enforce EPV ≥ 10 (events per variable), use nested cross-validation, and report external validation Why: Over-fitting produces signatures that appear strong internally but fail completely on any external cohort
  • Wrong: Choosing an imaging biomarker by convenience rather than biological relevance Right: Run the Decision Framework (question → modality → biomarker) — the biomarker should be mechanistically linked to the clinical question Why: Using whole-brain volume when the question is hippocampal atrophy, or ADC when the question is perfusion, measures the wrong thing
  • Wrong: Deciding QC exclusion thresholds after seeing the analysis results Right: Pre-specify QC metrics and exclusion thresholds before analysis; report inclusion/exclusion counts transparently Why: Post-hoc threshold selection biases the study toward desired results
  • Wrong: Smoothing or interpolating through a lesion or tumor ROI Right: Keep ROI-based measurements in native space, or use lesion-filling / cost-function masking before template normalization Why: Smoothing across lesion boundaries corrupts texture features and blurs lesion edges, invalidating radiomics and lesion-aware analyses
  • 错误： 将每个纵向时间点独立配准到 MNI 正确： 构建受试者内模板，或先将随访扫描刚性配准到基线；仅在需要时再归一化到 MNI 原因： 受试者特异性萎缩会被吸收到形变场中，纵向变化信号会丢失
  • 错误： 仅依赖基于标签的 DICOM 去标识化 正确： 在去标识化方案中纳入像素检查、私有标签策略、SR/PDF 处理、UID 重映射和 defacing 原因： 烧录在像素中的 PHI、私有标签、SR/PDF 叙述文本以及编码日期的 UID 都会在简单的标签级清理后残留
  • 错误： 使用 test-retest ICC < 0.75 的特征构建 radiomics 模型 正确： 在任何监督式选择之前，按 ICC ≥ 0.75（最好 ≥ 0.9）和分割稳健性筛选特征 原因： 不稳定特征会破坏可复现性——由其构建的特征签名无法推广至外部队列
  • 错误： 在多中心影像研究中忽略扫描仪或中心效应 正确： 标准化 protocol，采用 ComBat 或混合效应建模，将 site 作为协变量，并预注册分析计划 原因： 中心混杂可能完全解释表面上的生物学效应，从而产生假阳性发现
  • 错误： 对事件数较少的数据集拟合数十个 radiomics 特征 正确： 执行 EPV ≥ 10（events per variable），使用 nested cross-validation，并报告 external validation 原因： 过拟合产生的特征签名在内部验证中表现强劲，却会在任何外部队列中完全失效
  • 错误： 根据便利性而非生物学相关性选择影像生物标志物 正确： 执行决策框架（问题 → 模态 → 生物标志物）——生物标志物应在机制上与临床问题相关 原因： 当问题关注海马萎缩时使用全脑体积，或当问题关注灌注时使用 ADC，测量的其实是错误的对象
  • 错误： 在看到分析结果后才决定 QC 排除阈值 正确： 在分析前预先指定 QC 指标和排除阈值；透明报告纳入与排除数量 原因： 事后选择阈值会使研究偏向期望的结果
  • 错误： 对病灶或肿瘤 ROI 进行平滑或插值 正确： 将基于 ROI 的测量保留在原生空间，或在模板归一化前使用 lesion-filling / cost-function masking 原因： 跨越病灶边界进行平滑会破坏纹理特征并模糊病灶边缘，使 radiomics 和病灶感知分析失效

References

参考资料