shield_lock
金大哥
← 返回 Skill 列表
extension
分类: 效率与办公无需 API Key

medical-imaging-methods

Drafting and reviewing the Methods section of papers involving medical imaging, radiomics, or deep learning on imaging modalities (CT, MRI, X-ray, US, WSI). Covers acquisition reporting, preprocessing pipelines, segmentation methodology, feature extraction (PyRadiomics + IBSI), DL architectures (nnU-Net, TotalSegmentator), validation strategy, and metric selection (Dice, HD95, AUC, calibration, c-index). Use when writing or reviewing the Methods section of an imaging-based study.

person作者: TashanworldhubOpenAPI
open_in_new仓库download下载资源包

Medical Imaging Methods

When to use

  • Writing the Methods section of a paper involving CT/MRI/X-ray/US/WSI.
  • Reviewing such a section for compliance and reproducibility.
  • Adapting a thesis chapter on a radiomics/DL imaging pipeline into manuscript form.
  • Documenting a pipeline for internal protocols (CERT submission, study log).

This skill complements clinical-paper-writing — that one handles overall manuscript structure; this one handles the imaging-specific technical depth.

Process

Methods for an imaging study should cover, IN THIS ORDER:

1. Imaging acquisition

Reviewers will reject methods that don't report:

  • Modality and scanner manufacturer/model(s): e.g., "Siemens SOMATOM Force, GE Revolution CT".
  • Acquisition parameters: kVp, mAs (or CTDIvol), slice thickness, pixel spacing, reconstruction kernel (sharp vs soft tissue), iterative reconstruction strategy if used.
  • Contrast protocol (if applicable): agent, dose, injection rate, timing of acquisition.
  • Multi-vendor heterogeneity: if the cohort spans scanners, state this and how it was handled (harmonization or stratification).

For longitudinal data, also report:

  • Number of timepoints per patient.
  • Acquisition interval (median, IQR).

2. Preprocessing

Document EXACTLY:

  • Resampling: target voxel spacing (e.g., 1×1×1 mm) and interpolation method (linear for image, nearest for mask).
  • Intensity normalization: HU clipping range for CT (e.g., [-1024, 600] for lung), z-score / Nyul / WhiteStripe for MRI, why this choice.
  • Registration (if multi-sequence or longitudinal): rigid/affine/deformable, software (ANTs, Elastix, SimpleITK).
  • Cropping/ROI definition: how the bounding box was determined.
  • Anonymization: tool used (e.g., dcm2niix + custom script, CTP, dcm-anonymizer), what tags were modified, whether PixelData was inspected for burned-in PHI.

3. Segmentation (if applicable)

Choose ONE primary path and justify:

| Approach | When to use | What to report | |---|---|---| | Manual by experts | Gold standard, small cohort | Number of annotators, training, blinding, inter-/intra-observer ICC or DSC | | Semi-automatic | Pragmatic, need expert verification | Tool, default parameters, manual edit time, expert reviewer | | Fully automated (nnU-Net, TotalSegmentator) | Large cohort, validated tool | Version, model checkpoint, manual quality review protocol, exclusion criteria for failed cases |

Always report:

  • Reference/ground truth definition.
  • Quality control: whether segmentations were reviewed, by whom, blinded to outcome.
  • Exclusions due to segmentation failure (count and why).

4. Radiomic feature extraction (if applicable)

  • Software + version: e.g., PyRadiomics v3.1.0, MITK, LIFEx.
  • IBSI compliance: state explicitly whether features are IBSI-compliant. If using PyRadiomics, note that not all features are IBSI-validated by default.
  • Feature classes extracted: shape, first-order, GLCM, GLRLM, GLSZM, NGTDM, GLDM. State filters applied (LoG with which sigmas, wavelet which decomposition).
  • Discretization: bin width or bin count, justification.
  • Number of features extracted (raw, before selection).
  • Test-retest reproducibility: if available, ICC threshold for retention.
  • Harmonization: ComBat for multi-scanner data, with what covariates protected.

5. Deep learning model (if applicable)

  • Architecture: name + reference (nnU-Net, U-Net, ResNet-50, DINOv2, ViT). If custom, describe with a diagram.
  • Pretraining: from scratch, ImageNet, RadImageNet, MedicalNet, self-supervised on internal data.
  • Input format: 2D slices, 2.5D, 3D patches (size), full volume.
  • Training:
    • Optimizer (Adam/AdamW/SGD), learning rate, schedule.
    • Loss function (cross-entropy, Dice, focal, compound).
    • Batch size.
    • Epochs / steps and stopping criterion (early stopping with patience N on val metric).
    • Augmentation: list specifically (rotation, flip, elastic, intensity, MixUp, etc.).
    • Class imbalance handling (weighted sampling, focal loss).
  • Hardware: GPU model, count, framework + version (PyTorch 2.x), CUDA version. This is for reproducibility and runtime comparison.
  • Reproducibility: random seeds set, deterministic flags, container hash if Singularity/Apptainer/Docker.

6. Validation strategy

The single most scrutinized part. Be explicit:

  • Split level: patient-level. State how this was enforced (especially for longitudinal data where one patient has multiple scans).
  • Split ratios: e.g., 70/15/15 or k-fold cross-validation (k=5).
  • Internal vs external: if external, describe the external cohort separately with its own characteristics table.
  • Stratification: outcome, sex, age group, scanner.
  • Lock: state that the test set was held out and accessed only once for final reporting.

7. Metrics

Pick metrics matched to the task:

| Task | Primary | Secondary | Always report | |---|---|---|---| | Segmentation | Dice (DSC) | HD95, ASSD, sensitivity, specificity per voxel | Per-class breakdown if multi-class; volume per case | | Detection | Sensitivity at FPs/scan, FROC | Free-response AUC | Per-lesion size stratification | | Binary classification | AUC | F1, precision, recall, PR-AUC | Calibration (Brier, ICI), decision curve analysis (DCA) | | Multi-class | Macro AUC | Per-class AUC, confusion matrix | Class support | | Survival/time-to-event | C-index (Harrell or Uno) | Time-dependent AUC, IBS | Calibration (calibration plot per quantile), competing risks if relevant | | Prognostic biomarker | Hazard ratio with CI | Kaplan-Meier with log-rank | Number at risk under curves |

For ALL metrics:

  • 95% CI via bootstrap (state B = 1000 typically, percentile or BCa).
  • Comparison to a meaningful baseline (radiologist, clinical score, prior model).

8. Statistical comparisons

  • DeLong test for AUCs (independent), bootstrap for paired.
  • McNemar for paired sensitivity/specificity at fixed threshold.
  • Wilcoxon signed-rank for paired Dice.
  • State whether multiple-comparison correction was applied (Bonferroni, Holm, FDR).

Anti-patterns

| Pattern | Why it fails | Correction | |---|---|---| | "Standard preprocessing was applied" | Not reproducible | List every step with parameters | | Image-level split with patients having multiple scans | Test set leakage | Patient-level split, with code excerpt in supplementary | | Reporting only Dice for segmentation | Misses boundary errors | Add HD95 and ASSD; report per-class; show failure cases | | "Augmentation was used" without specifics | Reviewers will ask | Itemize: rotation ±10°, flip H, elastic with α=..., etc. | | PyRadiomics features called "IBSI-compliant" without verification | Some PyRadiomics features are NOT IBSI-validated | Cite IBSI list, state which features were retained as compliant | | AUC reported without CI | Single number is unfalsifiable | Bootstrap 95% CI mandatory | | No calibration for prediction models | Discrimination ≠ calibration | Report Brier score, calibration plot, ICI | | Multi-scanner cohort without harmonization | Confounded performance | Apply ComBat or report per-scanner stratified results | | GPU model not reported | Can't reproduce timing or replicate | State GPU + framework + CUDA versions |

Verification gates

Before declaring Methods done:

  • [ ] Every parameter mentioned has a value (no placeholders).
  • [ ] Software with versions (don't say "Python", say "Python 3.11").
  • [ ] Patient-level split confirmed in code; flow diagram shows n at each stage.
  • [ ] At least one metric with calibration assessment if it's a prediction task.
  • [ ] Comparator stated for primary metric.
  • [ ] Random seeds and reproducibility: documented.
  • [ ] Code/data availability statement drafted (even if "available on request").
  • [ ] If radiomics: IBSI compliance addressed.
  • [ ] If multi-scanner: harmonization addressed.
  • [ ] Quality control of segmentations described.

Notes on common mistakes specific to your projects

  • a CT-based prognostic model (competing risks): state that competing risks were modeled with Fine-Gray or cause-specific Cox; report cumulative incidence functions, not Kaplan-Meier as if it were single-cause; calibration via cause-specific calibration plots.
  • a lung cancer screening cohort nodule project: explicitly state Lung-RADS version used, and how borderline cases were handled.
  • a longitudinal post-COVID cohort longitudinal: patient-level split must respect longitudinal structure; state which timepoint(s) are inputs and which are outputs.
  • an ICU follow-up cohort subset of PostUCI: clearly state the relationship between cohorts to avoid reviewer confusion.