Medical Imaging Review

Medical Imaging AI Literature Review Skill (v3.0.0)

Produce comprehensive reviews that pass first-round peer review on factual grounds, not just structural grounds.

This is not a template-filling skill. It is a write-with-verify discipline.

---

Quick Start

A review project lives in 4 files (3 you write, 1 the skill provides):

project_root/
├── PARADIGM.md            # Style spec from 2-3 exemplar reviews (Phase 0)
├── CLAUDE.md              # Project-specific terminology + literature inventory
├── IMPLEMENTATION_PLAN.md # 3-axis outline + per-claim verification checklist
└── manuscript_draft.md    # The actual manuscript

Follow the 6-phase workflow in references/WORKFLOW.md. The phases are: paradigm capture → init → collect-and-verify → 3-axis outline → write-with-per-claim-verification → multi-agent peer review.

---

Core Principles

Writing voice — match strength to evidence, not hedge by default

Calibrate language to evidence strength, not to a fixed hedging register.

When ≥2 independent peer-reviewed groups confirm a finding, state it strongly. When evidence is single-source or contested, state it cautiously. When evidence is absent, say so.

Avoid the LLM tells:

• "has shown promising results"
• "may suggest"
• "interestingly,"
• "it is worth noting that"
• "in recent years,"
• "demonstrates the effectiveness of"
• "may offer significant advantages"

These phrases are AI-detector top features. Real flagship-review authors don't use them. Strip them.

Take a position when evidence supports it. Neutral catalogue is the LLM default and the failure mode to avoid. See Verdict sentences below.

Citations — every claim verified before commit

Every [N] citation must satisfy four checks:

1. The cited paper exists (DOI / PMID resolves on PubMed or Crossref). 2. The author list matches the first-source (especially first and last author). 3. The numeric claim in the body sentence (Dice, HR, sample size, etc.) appears in the cited paper's abstract or results section. 4. The directional claim in the body sentence (higher/lower, increased/decreased) matches the source's stated direction.

If any check fails, the citation is broken — fix before continuing. See references/CITATION_INTEGRITY.md for the full protocol.

Method descriptions — read first, write after

Do not fill in a template like [Author] et al. [ref] proposed [method]... Achieves Dice of X.XX. That template is a hallucination trap.

Use this discipline instead:

1. Read the actual paper (abstract + methods + results). For arXiv, use read_paper. For closed-access, use Zotero MCP to access the user's library. 2. Note the actual module names, the actual benchmark, the actual numbers, in your own working notes — not in the manuscript yet. 3. Write the method description from those notes, citing specific numbers and module names verbatim from the paper. 4. Verify by spot-checking 1-2 of the numbers against the paper one more time before moving on.

If you can't access the paper, do not write about its internal architecture or specific performance numbers. Cite it for the contribution-level claim only ("first to apply X to Y") and move on.

Heading depth — Nature Reviews uses two levels

• H2 (##) for top-level sections (Introduction, Methods, Applications, Discussion, ...).
• H3 (###) for subsections.
• H4 (####) is forbidden in body. Use bold lead-in **Topic.** paragraph starters for deeper grouping.
• Number prefixes (1., 1.1, 1.2.3) are forbidden in section titles. Nature Reviews / Nat Med / Lancet / JACC don't use them in narrative reviews.

Equations — in a Box, not in body

Display equations (DSC, IoU, clDice, FedAvg, GCN propagation, ...) appear in Boxes, not inline in body paragraphs. Textbook formulas can be referenced ("the Dice similarity coefficient — see Box 1") but should not be displayed inline.

If a formula has no methodological insight worth displaying (e.g., FedAvg averaging), describe it in prose instead of showing it.

Vendor names — only in the regulatory/products table

Vendor names (HeartFlow, Cleerly, Caristo, Keya, Shukun, ...) appear ONLY in the Commercial Products / Regulatory & Validation table. In body text use category descriptors:

• ✗ "HeartFlow's CT-FFR product was validated in NXT, ADVANCE, and PACIFIC..."
• ✓ "The first FDA-cleared CT-FFR product (Table N, row 1) was validated in NXT, ADVANCE, and PACIFIC..."

Reason: scatter-cited vendor names look like marketing copy and undermine the review's authority.

---

Standard Review Structure

# [Title]: <evocative subtitle>

## Key Points
- 4-5 bullets, each 1-3 sentences, expressing the main conclusions.

## Abstract

## Introduction
### Clinical background
### Technical challenge
### Scope and contributions

## Datasets and evaluation metrics
(Table 1: public datasets)
(Box 1: evaluation metrics with equations)

## Methods                              # 3-axis grouping, NOT flat 10-subsection list
### Architectural priors
**CNN-based design.** ... (bold lead-in for sub-grouping)
**Transformer-based design.** ...
**Mamba and state-space design.** ...

### Inductive priors
**Topology-aware design.** ...
**Multi-task design.** ...
**Graph-based design.** ...

### Data regime
**Self-supervised pre-training.** ...
**Foundation models.** ...
**Federated learning.** ...
**Physics-informed models.** ...

(Table 2: representative methods with modality / family / dataset / metric)

## Downstream applications
### [Application 1]
### [Application 2]
### [Application 3]

## Translation to clinical practice
(Table 3: commercial products with regulatory + validation)

## Outstanding challenges

## Future directions

## References

Notes:

• No number prefixes on headings.
• §Methods is 3 H3 subsections (the three axes), with bold lead-ins for each method family inside.
• Tables 1, 2, 3 are typically enough. Box 1 (metrics) is typical. Avoid 5+ tables.
• Verdict sentences cluster at the end of §Methods axis subsections and at the end of clinical translation discussions — not after every paragraph.

---

Verdict Sentences

Each H3 method-axis subsection (Architectural priors / Inductive priors / Data regime) should close with one verdict sentence expressing authorial position. Choose the 3-5 most opinionated positions across the whole manuscript — don't put verdicts on every paragraph.

Verdict templates:

• "[Family] is currently the most cost-effective design choice for [problem]."
• "[Family] has yet to demonstrate clear advantage over [alternative] in clinical-grade evaluations."
• "[Family] is best understood as complementary to [alternative], not a replacement."
• "The next [N] years will determine whether [family] becomes the default backbone or remains a research curiosity."

Neutral catalogue is the LLM default and exactly what flagship review editors push back on. Force yourself to take 3-5 positions.

---

Required Elements

• Key Points box (4-5 bullets, 1-3 sentences each) after the title.
• Tables 1-3: datasets, methods, commercial products.
• Box 1: evaluation metrics with formulas.
• Figures: typically 3-5 (overview/taxonomy, representative architectures, workflow, performance landscape).
• References: cite only what supports the argument. Quantity is downstream of substance — don't pad to a target count.
• Verdict sentences: 3-5 across the whole manuscript, clustered at axis-section ends.

---

Heading Depth

See Core Principles ▸ Heading depth above. Hard rules:

• Max 2 heading levels in body.
• No number prefixes.
• Use bold lead-in **Topic.** for deeper subsubsections.

---

Equations

See Core Principles ▸ Equations above. All display equations go in Box 1 (or rare additional Boxes for specific protocols). Textbook formulas with no methodological insight should be described in prose, not displayed.

---

Vendor Names

See Core Principles ▸ Vendor names above. Vendor names live in Table 3 only; body text uses category descriptors with table cross-reference.

---

Citation Style

# Data citation
"...achieved Dice of 0.730 on ImageCAS [N]"

# Method citation
"Xu et al. [N] introduced..."

# Multi-citation (max 4 in one bracket — beyond that, regroup the claim)
"Multiple groups demonstrated this effect [N1, N2, N3]"

# Comparative
"While [N1] focused on architecture, [N2] addressed the data side"

[N] in body must match the bibliography entry [N], and bibliography [N] must be the paper the body sentence is actually attributing the claim to. See references/CITATION_INTEGRITY.md Rule 3.

---

Literature Sources

Use all three in combination:

For closed-access journals (Med Image Anal, Eur Radiol, Lancet family) the user's local Zotero library is often the only path. Always check Zotero before assuming a paper is inaccessible.

For MCP server configuration, see references/MCP_SETUP.md.

---

Reference Files

---

Related Skills

For revising an existing AI-drafted review (whether your own previous output or someone else's draft), use ai-review-revision. That skill is the dedicated tool for fixing draft-quality issues — multi-agent diagnostic, factual reset, structural reset, content polish, submission prep.

This skill (medical-imaging-review) is the dedicated tool for producing draft-quality content correctly the first time. They are complementary:

• medical-imaging-review = write-side (produce submission-quality first draft)
• ai-review-revision = revise-side (rescue a draft that already has quality issues)

If a draft produced by this skill still ends up needing the ai-review-revision workflow to land, that's a bug — flag it so this skill can be improved.

---

Why this skill was rewritten (v3 vs v2)

v2.0.0 produced the coronary-cta-paper initial draft. That draft needed extensive multi-day revision before submission-readiness: 17 placeholder DOIs, 30-40 [N] citation drift errors, fabricated method module names with wrong performance numbers, vendor-style citations attributed to peer-reviewed journals, a 10-subsection flat method taxonomy where 3 thematic axes would have served better, AI-tone hedging language throughout.

v3 directly addresses each of these failure modes:

Citation Integrity Protocol

This is the most important reference file in the skill. Citation drift / fabrication is the #1 failure mode of LLM-drafted reviews.

5 rules. Every citation must satisfy all 5.

---

Rule 1: No placeholder DOIs

Any DOI containing xxx, [TBD], ?, stub patterns, or "to-be-filled" notation must be resolved before the manuscript advances to the next section.

Detection

grep -nE "xxx|\[TBD\]|x\):xxx|doi:10\.[a-z]+/x" manuscript_draft.md

Expected: 0 hits at any milestone.

Resolution

For papers with a PMID:

WebFetch on https://pubmed.ncbi.nlm.nih.gov/<PMID>/
  → Extract: title, full author list, journal, year, volume, issue, pages, DOI

For papers without a PMID but a known DOI:

WebFetch on https://api.crossref.org/works/<DOI>
  → Extract: same fields from JSON

For papers in a closed-access journal where the user has the PDF:

mcp__zotero__zotero_search_items(query: "<author> <method>")
mcp__zotero__zotero_get_item_metadata(item_key: ...)

Why this is Rule 1

v2 of this skill shipped 17 placeholder DOIs in the coronary-cta-paper draft. Editors at any medical journal will reject a manuscript with xxx in the DOI on first sight. This is the lowest-effort, highest-impact rule.

---

Rule 2: Author list verification

Every reference in the bibliography must have its first and last author verified against the first-source. For references with > 6 authors, list first 6 + et al. — but the first author (and last author for medical clinical papers) must be verbatim.

LLM failure mode

LLMs generate generic-sounding 4-author lists when they don't know the actual authors:

• ✗ "Liu Y, Zhang H, Chen X, Wang J. TransCC..." (these are common Chinese surnames in 4-author pattern — strong fabrication signal)
• ✗ "Smith J, Johnson A, Williams M, Brown D" (English equivalent)
• ✗ "Patel R, Kumar A, Singh P, Sharma N"

Detection

For each reference in the bibliography, ask:

• Are these 4 of the most common surnames in a major language?
• Do the initials look suspiciously "average" (J / A / M / D)?
• Is this a paper where I might have looked up the title but never the authors?

For any "yes" answer, treat as suspect and verify.

Resolution

WebFetch on https://pubmed.ncbi.nlm.nih.gov/<PMID>/
  OR
WebFetch on https://api.crossref.org/works/<DOI>
  → Replace the entire author list verbatim.

For arXiv papers:

WebFetch on https://arxiv.org/abs/<id>

Why this matters

A wrong author list reads as obvious fabrication to anyone in the field. Real reviewers know the principal investigators in their area, and seeing "Liu Y, Zhang H, Chen X, Wang J" attached to a high-profile method paper is an instant credibility kill.

---

Rule 3: Body ↔ Bibliography reconciliation

The [N] in the body must (a) exist in the bibliography and (b) be the paper the body sentence is actually attributing the claim to.

LLM failure mode — "citation drift"

When LLM reorganizes sections or merges paragraphs, it often forgets to renumber citations. The body sentence says "Shit et al. introduced clDice [43]" but bibliography entry [43] is a completely different paper (e.g., a centerline DRL paper). Bibliography entry [10] is actually the clDice paper.

Detection

# Find all citations in body
grep -nE "\[([0-9]+)\]" manuscript_draft.md | head -50

# Find bibliography entries
grep -nE "^[0-9]+\." manuscript_draft.md

For each body citation [N], verify:

• Is there a bibliography entry numbered N?
• Does that entry attribute to the right paper?

Resolution

Two patterns:

Pattern A: Single-author misattribution. Body says "Shit et al. [43]" but [43] is wrong. Action: grep -n "Shit" to find the correct number, edit body to point there.

Pattern B: Numerical drift across many citations. Body [60] should be [36] (after section reorganization shifted 24 entries). Action: list affected range, manually re-number.

For a typical 120-reference review, expect 5-15 drift instances if you're careful and 30-40 if you're not.

Detection at scale

Build a verification script:

# Pseudocode
for n in body_citations:
    bib_entry = bibliography.get(n)
    body_sentence = sentence_containing(n)
    body_author = extract_author_mention(body_sentence)
    bib_author = extract_first_author(bib_entry)
    if body_author and bib_author and body_author.lower() not in bib_author.lower():
        flag(n, body_author, bib_author)

---

Rule 4: Conclusion-direction verification

For every cited finding (HR, OR, p-value, effect size, "higher" / "lower" claim), the body sentence's directional claim must match the source's stated direction.

LLM failure mode

LLMs flip directions when paraphrasing. The paper says "patients with collateral circulation had lower FAI values." LLM paraphrase becomes: "Lv et al. [N] showed collateral circulation is associated with higher FAI."

This is one of the worst failures because the cited paper genuinely exists, the author is real, and the topic is right — but the conclusion direction is reversed. Catches eye of any reviewer in the area.

Detection

For every body sentence that contains directional language (higher / lower / increased / decreased / better / worse), the writer must have explicitly verified the direction against the source.

Quick checklist when writing such a sentence: 1. What direction does the source actually report? 2. What direction am I about to write? 3. Do those match?

Resolution

Always cite directions verbatim from the abstract. Don't paraphrase quantitative directional claims.

✗ "Lv et al. [N] showed collaterals are associated with higher FAI." ✓ "Patients with collateral circulation had lower FAI values than those without (Lv et al. [N])."

---

Rule 5: First-source over vendor materials

Vendor white papers, NHS England reports, FDA 510(k) clearance letters, and company press releases are not peer-reviewed primary sources for clinical findings. Cite them only for regulatory facts (clearance date, indication), not for clinical evidence (trial results, performance numbers).

LLM failure mode

LLM cites "NHS England. FISH&CHIPS Study Implementation Report. BMJ Open. 2024" as if it's a peer-reviewed trial. Reality: there is no such BMJ Open paper. The FISH&CHIPS study is a real NHS program, and its peer-reviewed publication is:

Fairbairn TA, Mullen L, Nicol E, Lip GYH, Schmitt M, Shaw M, et al. Implementation of a national AI technology program on cardiovascular outcomes and the health system. Nat Med. 2025;31(6):1903-1910.

The body sentence was right about the study existing, but the citation was a fabricated journal attribution.

Detection

For every reference that looks like:

• "[Company name]. [Product] White Paper. ..."
• "[Agency]. [Study]. ..."
• "[Study Investigators]. [Findings]. [Major journal]. 2024."

Treat as suspect and search for the actual peer-reviewed publication.

Resolution

For each vendor- or agency-style citation, search PubMed:

mcp__pubmed-mcp-server__pubmed_search_articles(query: "<study acronym> implementation OR validation")

If a peer-reviewed publication exists, use it. If not, use the vendor material only for regulatory / programmatic facts, never for clinical performance claims.

---

Verification Workflow Integration

These 5 rules are applied at multiple points:

---

What to do when a rule failure is found

Failures are normal — these are guardrails, not aspirational goals.

For each failure:

1. Stop forward writing immediately. Don't accumulate broken citations. 2. Look up the correct metadata using the resolution steps above. 3. Fix the citation in place — body sentence + bibliography entry both. 4. Check for related failures. A misattribution in one place often signals the same error in 2-3 other places where the same paper was cited. 5. Log the fix in IMPLEMENTATION_PLAN.md change log — this protects the next person editing the manuscript.

---

Why this protocol exists

The single largest source of credibility damage in LLM-drafted reviews is citation infrastructure failure. A draft can have brilliant analysis and miss editor first-glance if the bibliography is fabricated, drifted, or misattributed.

This protocol takes 10-20% of total writing time but eliminates 80%+ of the credibility-killer issues. It is non-negotiable.

Domain-Specific Method Taxonomies (3-Axis Grouping)

This file replaces the v2 "flat 10-subsection list" approach. For each medical imaging domain, methods are organized along 3 thematic axes, not as a long flat list. These axes are subsection structure (the H3 organization in §Methods), with bold lead-ins for individual method families inside each axis.

The 3 axes are universal: 1. Architectural priors — what kind of network (CNN, Transformer, Mamba, etc.) 2. Inductive priors — what kind of geometric / structural / multi-task bias is built in (topology, multi-task, graph, etc.) 3. Data regime — how data is used / pre-trained / federated (self-supervised, foundation models, federated, etc.)

Critical: These 3 axes are how the §Methods section is structured. Do NOT bullet 10 method categories as 10 H3 subsections. Use 3 axis subsections, with bold lead-ins inside each axis for the method families.

---

How to use this file

When writing the §Methods section:

1. Open this file to find your domain. 2. Use the 3-axis grouping as your H3 structure. 3. Inside each axis subsection, use bold lead-ins for the individual method families. 4. End each axis subsection with a verdict sentence.

Example for coronary segmentation §Methods:

## Methods

### Architectural priors

**CNN-based design.** ... 2-3 paragraphs ...
**Transformer-based design.** ... 2-3 paragraphs ...
**Mamba and state-space design.** ... 1-2 paragraphs ...

Verdict: CNN-based design remains the operational default; transformer
hybrids are starting to show convincing gains in centerline-vs-mask hybrid
problems but have yet to displace U-Net on pure segmentation.

### Inductive priors

**Topology-aware design.** ... 2-3 paragraphs ...
**Multi-task design.** ... 2-3 paragraphs ...
**Graph neural network design.** ... 1-2 paragraphs ...

Verdict: Topology-aware losses are the single most cost-effective design
choice for coronary segmentation when paired with any decent backbone.

### Data regime

**Self-supervised pre-training.** ... 2-3 paragraphs ...
**Foundation models.** ... 2-3 paragraphs ...
**Federated learning.** ... 1-2 paragraphs ...
**Physics-informed models.** ... 1 paragraph ...

Verdict: Foundation models are the next 2-3 years' wild card; their gap to
domain-tuned specialists has narrowed substantially but is not yet closed.

---

Coronary Artery / Cardiovascular CT (CCTA)

Axis 1: Architectural priors

• CNN-based: U-Net, V-Net, nnU-Net variants
• Transformer-based: ViT, SwinUNETR, TransUNet, TransCC, FocusUNETR
• Mamba / state-space: VM-UNet, U-Mamba for vessels
• Hybrid CNN-Transformer: nnFormer, hybrid encoders

Axis 2: Inductive priors

• Topology-aware: clDice loss, VCP loss, persistent-homology losses
• Multi-task: joint segmentation + centerline, joint segmentation + bifurcation detection
• Graph neural network: vessel graph extraction, GNN-based labeling

Axis 3: Data regime

• Self-supervised pre-training: contrastive, masked-image-modeling for vessels
• Foundation models: SAM-Med, vesselFM, generalist segmentation models
• Semi-supervised: pseudo-labeling for CCTA datasets
• Federated learning: multi-center coronary segmentation without data sharing
• Physics-informed: PDE-constrained losses for vessel topology

Downstream tasks (separate section, not part of methods axis structure)

• Centerline extraction
• Vessel labeling (SCCT 18-segment / AHA 17-segment myocardium)
• Stenosis detection
• CT-FFR computation
• Plaque analysis
• Calcium scoring
• Pericoronary fat analysis (FAI)

Key datasets

• CAT08 (32 cases, centerline)
• ASOCA (40 cases, segmentation)
• ImageCAS (1000 cases, single-center, segmentation)
• PCCTA120 (120 cases, artery + plaque)

---

Lung Imaging (CT / X-ray)

Axis 1: Architectural priors

• Anchor-based detection: Faster R-CNN, RetinaNet
• Anchor-free detection: CenterNet, FCOS, YOLO variants
• Transformer-based detection: DETR family
• 3D detection: 3D nodule detection networks
• U-Net variants for segmentation

Axis 2: Inductive priors

• Multi-scale feature pyramids: FPN-based, PSP
• Attention mechanisms: SE blocks, CBAM, axial attention
• Boundary-aware: edge-loss formulations
• Uncertainty quantification: MC dropout, ensembles

Axis 3: Data regime

• Self-supervised: contrastive learning on chest CT
• Weakly-supervised: from radiology reports
• Foundation models: chest X-ray foundation models

Tasks

• Nodule detection / segmentation / malignancy classification
• COVID-19 detection
• Interstitial lung disease characterization

Key datasets

• LUNA16 (888 CT scans)
• LIDC-IDRI (1018 cases)
• ChestX-ray14 (112,120 X-rays)

---

Brain Imaging (MRI / CT)

Axis 1: Architectural priors

• CNN-based: U-Net, V-Net
• Transformer-based: SwinUNETR, UNETR for BraTS
• Hybrid

Axis 2: Inductive priors

• Attention mechanisms: spatial, channel, self-attention
• Graph neural networks: brain connectivity GNNs
• Multi-atlas-informed: deep atlas registration

Axis 3: Data regime

• Self-supervised pre-training: masked-image-modeling on MRI
• Foundation models: medSAM, BrainSAM
• Multi-modal fusion: T1/T2/FLAIR fusion strategies
• Federated learning: cross-institutional MRI federation

Tasks

• Brain tissue segmentation
• Tumor segmentation (BraTS)
• Lesion detection (stroke, MS)
• Cerebrovascular segmentation
• Age / disease estimation

Key datasets

• BraTS (brain tumor)
• ADNI (Alzheimer's)
• IXI (healthy brains)
• ISLES (stroke lesions)

---

Cardiac Imaging (MRI / CT / Echo)

Axis 1: Architectural priors

• CNN-based: nnU-Net, V-Net cine MRI
• Temporal modeling: RNN, 3D CNN, transformer-based temporal
• Multi-view fusion: SA + LA cine fusion

Axis 2: Inductive priors

• Shape priors: SSM-constrained networks
• Anatomical loss formulations
• Uncertainty estimation: ensembles, MC dropout

Axis 3: Data regime

• Multi-modal fusion: cine + LGE + perfusion
• Foundation models: cardiac generalists
• Self-supervised pre-training

Tasks

• Chamber segmentation
• Wall motion analysis
• Scar / fibrosis detection (LGE)
• Valve assessment
• Strain analysis

Key datasets

• ACDC (100 patients)
• M&Ms (320 subjects)
• CAMUS (500 patients, echo)

---

Pathology (Whole Slide Images)

Axis 1: Architectural priors

• Patch-based CNN: ResNet, EfficientNet
• Transformer-based: ViT, hierarchical transformers
• Graph neural networks: nuclei-level graphs

Axis 2: Inductive priors

• Multiple Instance Learning (MIL): attention-MIL, max-pooling
• Attention-based aggregation: TransMIL
• Topology-aware: persistent homology of histological structures

Axis 3: Data regime

• Self-supervised pre-training: SimCLR / MoCo / DINO on patches
• Foundation models: PathLM, CONCH, UNI, Virchow
• Weakly-supervised: from slide-level labels

Tasks

• Cancer detection / grading / staging
• Biomarker prediction
• Survival prediction

Key datasets

• CAMELYON (lymph node)
• TCGA (multi-cancer)
• PANDA (prostate)

---

Retinal Imaging (Fundus / OCT)

Axis 1: Architectural priors

• CNN-based: multi-scale networks
• Transformer-based: ViT for fundus
• Hybrid

Axis 2: Inductive priors

• Attention mechanisms: dual-attention for vessels
• Domain adaptation: between fundus camera types

Axis 3: Data regime

• Self-supervised: on large unlabeled fundus image sets
• Foundation models: RETFound, FLAIR
• Federated learning: privacy-preserving DR screening

Tasks

• Diabetic retinopathy grading
• Glaucoma detection
• Age-related macular degeneration
• Vessel segmentation

Key datasets

• EyePACS (88,702 images)
• DRIVE (40 images, vessels)
• REFUGE (1200 images, glaucoma)

---

Universal Medical Image Segmentation (Fallback Axis Structure)

When the domain is generic or your topic spans multiple modalities, use this universal 3-axis grouping:

Axis 1: Architectural priors

• Encoder-Decoder (U-Net, V-Net, nnU-Net)
• Transformer-based (SwinUNETR, UNETR, TransUNet)
• Mamba / state-space
• Hybrid CNN-Transformer

Axis 2: Inductive priors

• Attention mechanisms (SE, CBAM, axial, deformable)
• Multi-scale processing (FPN, PSP, ASPP)
• Boundary-aware (active contours, edge losses)
• Topology-preserving (clDice, persistent homology)
• Uncertainty quantification (MC Dropout, ensembles)

Axis 3: Data regime

• Self-supervised pre-training (contrastive, masked)
• Foundation models (SAM, MedSAM)
• Few-shot / zero-shot (prototypical, foundation models)
• Domain adaptation (adversarial, self-training)
• Federated learning
• Efficient architectures (MobileNet, EfficientNet, Mamba — when efficiency is the focus)

Universal evaluation metrics (Box 1 content)

• Overlap: Dice, IoU / Jaccard
• Distance: Hausdorff (HD, HD95), ASSD
• Topology: clDice, Betti numbers
• Clinical: Sensitivity, Specificity, AUC, PPV / NPV

---

What to do when a method family doesn't fit cleanly into 3 axes

Some methods span axes (e.g., a foundation-model-based topology-aware Mamba would touch all 3). In such cases:

• Place the method in the axis it's most centrally about.
• Cross-reference from the other 2 axes ("see also: this method type combines architectural and data-regime innovations").
• Don't create a 4th axis to accommodate it. Three axes is the structural commitment.

If 30%+ of your methods don't fit, you may be in a sub-domain where the 3-axis structure needs adaptation. In that case, document the alternative structure in PARADIGM.md and use it consistently — but stay disciplined to 3 axes.

---

Why 3 axes, not 10 flat subsections

The v2 skill's coronary section listed 10 flat method categories. The resulting draft had a §3 with 10 nearly-equal H3 subsections, each ~500 words. The effect on the reader: a textbook chapter, not a flagship review.

Flagship reviews compress 10+ method variants into 3 thematic axes. The 3-axis structure also forces explicit comparison ("CNN-based vs Transformer-based vs Mamba-based architectures all aim to capture spatial inductive bias differently") which is what a real review reader wants — synthesis, not catalogue.

If your editor / reviewer feedback says "the methods section reads as a flat list", this file's structure is the fix.

9 Patterns of LLM Hallucination in Medical Imaging Reviews

Use this as a checklist during Phase 4 (writing) and Phase 5 (peer review). Every paragraph should be self-scanned against these patterns. The 9 patterns are derived from the actual failure modes observed in the coronary-cta-paper v2 draft.

Examples are drawn from coronary imaging but the patterns are domain-general.

---

Pattern 1: Real paper, fabricated author list

Most common. The paper exists, but the author list is a generic 4-name pattern (or only includes 1-2 real authors mixed with fabricated co-authors).

Detection signals

• Author list looks like "Liu Y, Zhang H, Chen X, Wang J" — 4 of the most common surnames in a language paired with most common initials.
• First-author initial doesn't match real first author (e.g., "Wittmann F" vs. real Bastian Wittmann).
• Author count obviously wrong for the journal (Nature paper with 3 authors is unusual; methods paper with 12 authors is unusual).
• Author list looks suspiciously "alphabetical" or "rhythmic" — real lists are messy.

Example from coronary-cta-paper

• v2 reference [30]: "Liu Y, Zhang H, Chen X, Wang J. TransCC..."
• Reality: Chenchu Xu, Meng Li, Xue Wu (3 actual authors of the real TransCC paper)

Fix

Replace entire author list with verbatim from arXiv / PubMed / Crossref / Zotero.

---

Pattern 2: Inflated performance numbers

The cited paper is real, the topic is right, but the specific Dice / sensitivity / specificity / HR is fabricated and typically inflated.

Detection signals

• Dice values ending in .891, .917, .943 — suspiciously specific 3-decimal numbers
• Performance numbers consistently above field benchmarks
• Round-looking numbers when papers usually report messier values
• Multiple methods all reporting "Dice 0.89" when the real distribution should span 0.70-0.85

Example from coronary-cta-paper

• v2: "TransCC achieves Dice 0.891 / HD95 1.72 mm on 120 cases"
• Reality: Dice 0.730 / IoU 0.582 on ImageCAS (not "120 cases")

Fix

Open the actual paper. Quote the actual numbers from abstract or results table. If the paper isn't accessible, drop the specific numeric claim and only cite the contribution level ("first to apply X to Y").

---

Pattern 3: Conclusion-direction flip

Most pernicious. Real paper, real authors, real finding — but the direction (higher/lower, increased/decreased) is reversed.

Detection signals

• Look for any body sentence with directional language: "higher", "lower", "increased", "decreased", "better", "worse", "more", "less"
• These all need explicit verification against the source

Example from coronary-cta-paper

• v2: "Lv et al. showed collateral circulation is associated with higher FAI"
• Reality: collaterals associated with lower FAI

Fix

Always quote directional claims verbatim from the source abstract. Don't paraphrase quantitative directional language.

---

Pattern 4: Vendor / agency material cited as peer-reviewed

Vendor white papers, FDA clearance letters, NHS reports cited with fabricated journal attribution.

Detection signals

• References that look like: "Company X. Product Y Whitepaper. Major Journal. 2024"
• "Study Investigators. Findings. BMJ Open. 2024" (when no such BMJ Open paper exists)
• Citations of clinical findings backed by a regulatory document rather than a peer-reviewed trial

Example from coronary-cta-paper

• v2: "NHS England. FISH&CHIPS Study Implementation Report. BMJ Open. 2024"
• Reality: real publication is Fairbairn TA, et al. Nat Med. 2025;31(6):1903-1910

Fix

Search PubMed for the actual peer-reviewed publication of the study. Use vendor materials only for regulatory / programmatic facts (clearance dates, indications), never for clinical claims.

---

Pattern 5: Placeholder DOIs

xxx, [TBD], x):xxx-xxx stubs left in the bibliography.

Detection

grep -nE "xxx|\[TBD\]|x\):xxx|doi:10\.[a-z]+/x" manuscript_draft.md

Example from coronary-cta-paper

• v2 had 17 entries like: Eur Radiol. 2026;36(x):xxx-xxx. doi:10.1007/s00330-025-xxx
• Reality: each has a real DOI resolvable via PubMed by PMID

Fix

For each placeholder, WebFetch on PubMed by PMID to extract real metadata. If the paper genuinely doesn't have a DOI yet (true preprint), use the arXiv ID or accept the "online ahead of print" notation — but never xxx.

---

Pattern 6: Generic 4-author hallucination (subset of pattern 1, called out separately)

Specifically the "4 common surnames + common initials" pattern. So distinctive it deserves its own watch.

Examples

• "Zhang H, Wang L, Chen Y, Liu Q" (Chinese)
• "Smith J, Johnson A, Williams M, Brown D" (English)
• "Patel R, Kumar A, Singh P, Sharma N" (Indian)
• "Sato K, Tanaka T, Suzuki M, Watanabe H" (Japanese)

Fix

Any time you see exactly 4 authors with this pattern, verify all 4 against the source. Real 4-author papers usually have at least one less-common name.

---

Pattern 7: Citation number drift

Body says "[43]" but bibliography [43] is the wrong paper. The correct paper is at [10] (or wherever).

Detection

For every paragraph, check 1-2 random [N] against bibliography:

• Is the body sentence's topic congruent with bibliography [N]'s title?
• Is the body sentence's author name (if mentioned) the bibliography [N] author?

Example from coronary-cta-paper

• v2: "Shit et al. [43] introduced clDice"
• v2 bibliography [43]: a centerline DRL paper (unrelated)
• v2 bibliography [10]: Shit S, Paetzold JC. clDice — CVPR 2021 (correct)
• Total drift instances in v2: 30-40

Fix

For each affected [N], grep for the correct number and edit body in place.

---

Pattern 8: Metric formula errors

Definitions of standard metrics are wrong. Often: clDice written as "sum-divided-by", actually defined as "harmonic mean".

Common confusions

• clDice: actually harmonic mean of topological precision and recall (HM(Tprec, Tsens)), often mis-written as "sum / 2" or "average"
• Hausdorff distance: actually max(d(A→B), d(B→A)), often mis-written as "average distance"
• IoU vs Dice: |A∩B|/|A∪B| vs 2|A∩B|/(|A|+|B|) — sometimes swapped
• FedAvg: weighted by client-data-size, often mis-described as simple averaging

Fix

For every metric or method formula displayed, verify against the original paper. Don't rely on memory — these are textbook formulas with subtle precise definitions.

---

Pattern 9: Internal inconsistency across sections

The same paper is cited in §3.2 with one author list and in §3.5 with a different author list. Or the same dataset appears with different patient counts in different sections.

Detection

Pick the 5 most-cited papers in the manuscript. For each, grep for every appearance:

grep -n "TransCC\|Wittmann\|FISH&CHIPS\|ORFAN\|<key-paper-name>" manuscript_draft.md

Check: every appearance of the same paper should attribute to the same author list with the same numbers.

Example from coronary-cta-paper

• v2: ImageCAS described as "1000 cases" in §2, "multi-center 1500 cases" in §3.2, "single-center 1000 cases" in Table 3
• Reality: ImageCAS is single-center, 1000 cases

Fix

For each high-value paper or dataset, lock in a canonical description in CLAUDE.md terminology table. All sections must conform.

---

Self-Check Workflow

Every 5-6 paragraphs during writing:

1. Scan the last block for direction language → verify pattern 3. 2. Look at any new citations → verify patterns 1, 2, 5, 6. 3. Cross-check any vendor/agency references → verify pattern 4. 4. Random-sample 1-2 [N] body↔bib → verify pattern 7. 5. Check any metric formula displayed → verify pattern 8.

Every section completion:

6. Search the section for the manuscript's top 5 cited papers → verify pattern 9 (internal consistency).

This adds ~10-15% to writing time. Catches >90% of hallucinations before they leave the writer's desk.

MCP Server Configuration for Literature Collection

Set up arxiv-mcp / pubmedmcp / zotero-mcp for use in Phase 2 (Literature Collection + Verification). For each server, this file documents installation, tools available, and the searches that match the workflow.

---

ArXiv MCP (Preprints & Latest Research)

Repository: https://github.com/blazickjp/arxiv-mcp-server

Configuration

Add to ~/.claude/mcp.json (or your MCP config file):

{
  "mcpServers": {
    "arxiv": {
      "command": "uvx",
      "args": ["arxiv-mcp-server"],
      "env": {
        "ARXIV_STORAGE_PATH": "~/.arxiv-mcp-server/papers"
      }
    }
  }
}

Available tools

Search strategy

Query: "[topic] AND (segmentation OR detection OR classification)"
Categories: cs.CV, eess.IV, cs.LG
Date: Last 3 years for current state of the art
Max results: 50-80 per query (discriminate aggressively — quality over breadth)

Example queries

• "medical image segmentation transformer" (cs.CV, eess.IV)
• "coronary artery deep learning" (cs.CV)
• "CT scan neural network" (eess.IV)
• "foundation model medical segmentation" (cs.CV, cs.LG)

Workflow integration

Per CITATION_INTEGRITY.md Rule 2, when adding an arXiv paper to bibliography:

1. search_papers to find candidates 2. download_paper(paper_id) for promising ones 3. read_paper(paper_id) to read full text (or at least abstract + methods + results) 4. Note actual first author, full author list, exact module names, headline numbers 5. Cross-check: arxiv abstract page = https://arxiv.org/abs/<id> for author list verification

---

PubMed MCP (Biomedical Literature)

Repository: https://github.com/grll/pubmedmcp

Access 35+ million biomedical literature citations.

Configuration

{
  "mcpServers": {
    "pubmedmcp": {
      "command": "uvx",
      "args": ["pubmedmcp@latest"],
      "env": {
        "UV_PRERELEASE": "allow",
        "UV_PYTHON": "3.12"
      }
    }
  }
}

Available tools

Search tips

• Use MeSH terms for precise medical searches
• Combine with publication type filters (Review, Clinical Trial)
• Filter by date for recent literature

Example MeSH queries

• "Deep Learning"[MeSH] AND "Coronary Vessels"[MeSH]
• "Image Processing, Computer-Assisted"[MeSH] AND "Tomography, X-Ray Computed"[MeSH]
• "Cardiac Imaging Techniques"[MeSH] AND "Artificial Intelligence"[MeSH]

Direct WebFetch for verification

pubmed_search_articles returns PMIDs. For metadata verification (CITATION_INTEGRITY.md Rules 1-2):

WebFetch on https://pubmed.ncbi.nlm.nih.gov/<PMID>/
  → Extract: full author list, journal, year, vol, issue, pages, DOI, finding direction

This is the canonical first-source verification step for medical clinical papers.

---

Zotero Integration

Access user's local Zotero database via Zotero-MCP.

Direct API Access (fallback)

# List collections
curl -s "http://localhost:23119/api/users/[USER_ID]/collections"

# Get items from a collection
curl -s "http://localhost:23119/api/users/[USER_ID]/collections/[KEY]/items"

Zotero-MCP (recommended)

Repository: https://github.com/54yyyu/zotero-mcp

Provides structured access:

Workflow integration

For closed-access journals (Med Image Anal, Eur Radiol, JACC family, Lancet family, Nature family), the user often has PDFs in Zotero that aren't accessible via WebFetch. Workflow:

1. mcp__zotero__zotero_search_items(query: "<author> <method>", limit: 5)
2. mcp__zotero__zotero_get_item_metadata(item_key: "<key>")
3. mcp__zotero__zotero_get_item_fulltext(item_key: "<key>") for body content

Extractable fields

• title
• abstractNote
• date
• creators (author list — verify against first-source per Rule 2)
• publicationTitle
• DOI
• tags
• collections

---

Source Selection Guide

---

Verification helper commands

For Phase 4 (per-claim verification) and Phase 5 (peer review):

# Crossref by DOI
WebFetch on https://api.crossref.org/works/<DOI>
  → Returns JSON: title, full author list, container-title (journal), volume, issue, page, DOI, published year

# Crossref by topic search
WebFetch on https://api.crossref.org/works?query.bibliographic=<keywords>&rows=5
  → Returns top 5 matching entries

# PubMed by PMID
WebFetch on https://pubmed.ncbi.nlm.nih.gov/<PMID>/
  → Returns parsed page: title, authors, journal info, DOI, abstract

# arXiv abstract page (for author list verification)
WebFetch on https://arxiv.org/abs/<id>
  → Returns abstract + full author list

---

When MCP servers fail

If an MCP server is not configured or fails:

• ArXiv fallback: WebFetch on https://arxiv.org/abs/<id> directly
• PubMed fallback: WebFetch on https://pubmed.ncbi.nlm.nih.gov/<PMID>/ directly
• Zotero fallback: ask the user to share PDFs directly, or use direct API access via curl

The skill is designed to work even without MCP — WebFetch on the underlying APIs is always available.

Phase 0: Paradigm Capture

Before writing the manuscript, extract a style spec from 2-3 exemplar reviews in the target journal tier. This anchors the draft to flagship-review style and prevents drift toward the generic "survey paper" register that LLMs default to.

---

Why this phase exists

LLMs trained on the entire web have absorbed every register of medical writing — vendor white papers, methodological surveys, textbook chapters, narrative reviews, systematic reviews, flagship reviews — and have no built-in preference among them. By default, the register they generate is somewhere between "good survey paper" and "Wikipedia article": numbered chapters, hedging language, neutral catalogue, dense subsections.

Flagship-tier reviews (Nature Reviews / Nat Med / Lancet family / JACC) write differently:

• 2 heading levels, no number prefixes
• 1-2 sentence "verdict" closers per axis section
• Equations and dataset details in Boxes, not body
• Vendor names confined to a single Table
• Authorial position taken explicitly when evidence supports it
• 1.5-2.5 references per paragraph, well-targeted
• Often a 4-5 bullet "Key Points" box after the title

Without an exemplar anchor, even careful writing drifts toward the generic register. With one, every paragraph has a benchmark.

---

Action 1: Select 2-3 exemplar reviews

Selection criteria

• Same modality or same problem family as your topic. (For coronary AI, look at recent coronary or cardiovascular AI reviews. For brain tumor segmentation, look at neuroradiology AI reviews.)
• Same journal tier as your target. If aiming for Nat Med, read a recent Nat Med review. If aiming for Eur Radiol, read a recent Eur Radiol review.
• Last 3 years. Older exemplars may have outdated stylistic conventions.
• Recognized authors. Reviews authored by senior figures in the field carry the conventions of their target journals more reliably.

Suggested exemplars by tier

Top tier (Nature Reviews / Nat Med / Lancet family):

• Nature Reviews Cardiology — for cardiovascular AI
• Nat Med — for AI-in-medicine narrative reviews
• Lancet Digital Health — for AI clinical translation

Second tier (specialty journals):

• JACC: Cardiovascular Imaging — for cardiac imaging AI
• Radiology / Eur Radiol — for general imaging AI
• IEEE TMI — for methodological reviews with clinical context

Methodological tier:

• Medical Image Analysis — for purely methodological reviews
• IEEE JBHI — for AI-in-healthcare methods

Acquiring the PDFs

Most flagship reviews are accessible via:

• The user's Zotero library (check first — they likely already have them)
• PubMed Central (for open-access)
• The journal's website (subscription / institutional access)
• arXiv preprint version (sometimes available)

---

Action 2: Read carefully — not skim

Spend 60-90 minutes per exemplar. Read the full text, including Boxes, Figure captions, and Table footnotes.

Pay attention to:

Heading structure

• How many levels deep? (Almost always 2 for top tier.)
• Are headings numbered or unnumbered? (Almost always unnumbered for top tier.)
• How long are H2 sections? (Typically 800-1500 words.)
• How long are H3 sections? (Typically 300-800 words.)
• How are deeper subsections handled? (Bold lead-in **Topic.**, not H4.)

Paragraph rhythm

• How does a paragraph open? (Topic claim — almost never "In recent years," or "It is worth noting.")
• How does evidence flow? (1-3 sentences of specifics with citations, not loose generalities.)
• How does a paragraph close? (Either with a transition or with a verdict — not with hedging.)
• How many sentences per paragraph? (Usually 4-7.)

Citation density

• Count references per paragraph in 5 random paragraphs.
• Top tier: usually 1.5-2.5 refs per paragraph.
• Lower density suggests overclaiming; higher suggests stacking refs uncritically.

Equation handling

• Are display equations in body or Boxes?
• If in body, is there a Box at all?
• How are metric definitions handled — formula or prose?

Vendor handling

• Search for any commercial product name. Does it appear in body text or only in Tables?
• How are FDA / regulatory facts cited?

Authorial voice

• Find the 3-5 strongest claims in the review. How are they phrased?
• Does the author take positions like "X is currently the best approach for Y" or stay neutral throughout?
• Where does hedging appear? (Usually only when evidence genuinely supports caution.)

Boxes and Figures

• How many Boxes? (1-3 typical.)
• What goes in a Box? (Definitions, key trials, controversies.)
• How many Figures? (3-5 typical.)
• What kinds of Figures? (Overview / taxonomy / workflow / data-driven plot.)

Tables

• How many Tables? (2-4 typical — rarely 5+.)
• What goes in a Table? (Datasets, methods comparison, commercial products.)
• How dense? (10-20 rows is common; 30+ is unusual.)

Key Points box

• Is there one? (Yes for most top tier.)
• How many bullets? (4-5 typical.)
• How long is each bullet? (1-3 sentences.)

---

Action 3: Write PARADIGM.md in the project root

Use this template:

# Paradigm Spec for [Project Name]

## Target journal: [name and tier]

## Exemplars studied
1. [Author Year]. [Title]. [Journal]. [PMID/DOI]. (PDF: <path>)
2. [Author Year]. [Title]. [Journal]. [PMID/DOI]. (PDF: <path>)

## Extracted style spec

### Heading structure
- Max depth: [N] levels
- Numbering: [yes / no]
- Typical H2 length: [N-N] words
- Deeper subsections: [H4 / bold lead-in / other]

### Paragraph rhythm
- Opening: [observed pattern, e.g., "Topic claim sentence, no temporal preamble"]
- Body: [observed pattern, e.g., "2-4 sentences of specifics with citations"]
- Closing: [observed pattern, e.g., "Transition or verdict"]
- Typical length: [N-N] sentences

### Citation density
- Average refs per paragraph: [N]
- Multi-citation cap (max refs in one bracket): [N]

### Equations
- Location: [body / Box / both / none]
- Display style: [LaTeX / inline / prose]

### Vendor handling
- Body: [allowed / disallowed]
- Table: [yes / no]
- Total Tables containing vendor names: [N]

### Authorial voice
- Hedging frequency: [observed examples]
- Strong claims: [N observed across 3 reviews]
- Verdict closers per section: [observed pattern]

### Boxes and Figures
- Box count: [N]
- Box purposes: [list]
- Figure count: [N]
- Figure types: [list]

### Tables
- Count: [N]
- Density: [N-N rows]

### Key Points
- Present: [yes / no]
- Bullet count: [N]
- Bullet length: [N-N sentences]

## Manuscript style targets (binding for Phase 4 writing)

Phase 4 writing should conform to this spec. Specifically:
- Max heading depth: ...
- No numbered headings.
- Box 1 for metrics / formulas.
- Vendor names only in Table N.
- ≥ 3 verdict sentences across the manuscript.
- Citation density 1.5-2.5 refs/paragraph.
- Key Points: 4-5 bullets after title.

## Anti-patterns observed in v2 drafting (to specifically avoid)

(Copy / adapt from this skill's HALLUCINATION_PATTERNS.md to the extent the patterns are domain-relevant.)

---

Action 4: Re-read PARADIGM.md before each Phase 4 writing session

Phase 4 is multi-day. Drift is the main risk. At the start of each writing session, re-read PARADIGM.md to re-anchor.

---

Time budget

• Exemplar selection: 30 min
• Reading 2 exemplars: 2-3 hours
• Writing the spec: 30-45 min

Total: 3-4 hours.

This is the highest-ROI 3-4 hours in the entire project. Skipping it almost guarantees survey-paper-register drift, which is then extremely expensive to fix in revision.

Quality Checklist for Medical Imaging AI Reviews

This checklist runs at multiple checkpoints — end of each section, end of each writing day, before peer-review phase, before submission.

v3 changes from v2: Dropped "hedging language used", "80-120 references", and format-only metric checks. Replaced with substance checks (citation integrity, conclusion direction, verdict presence, structural discipline).

---

Citation Integrity (Hard Gate)

These are non-negotiable. Any failure must be fixed before continuing.

• [ ] No placeholder DOIs (grep xxx, [TBD], x):xxx)
• [ ] Every reference's first and last author verified against first-source
• [ ] Body↔bibliography [N] reconciliation — spot-check 10 random citations per section
• [ ] Every directional claim (higher/lower, increased/decreased) verified against source
• [ ] No vendor white papers cited as if peer-reviewed
• [ ] No duplicate references (same paper listed under two numbers)

See CITATION_INTEGRITY.md for the full 5-rule protocol.

---

Structural Discipline

• [ ] Heading depth ≤ 2 levels (H2 + H3 only in body)
• [ ] No numbered headings (no ## 1., ### 1.1)
• [ ] H4 (####) absent from body — deeper grouping via bold lead-in **Topic.**
• [ ] §Methods uses 3-axis grouping (Architectural / Inductive / Data regime), not flat 10-subsection list
• [ ] Verdict sentences present in 3-5 places (not on every paragraph, not absent entirely)
• [ ] Key Points box has 4-5 bullets, 1-3 sentences each
• [ ] Standard sections present: Introduction / Methods / Applications / Translation / Discussion (names may vary, structure fixed)

---

Voice and Register

• [ ] No "has shown promising results" / "may suggest" / "interestingly" / "in recent years," / "it is worth noting" anywhere
• [ ] Hedging used only when evidence genuinely supports caution, not as default register
• [ ] Strong findings (≥2 independent groups confirming) stated strongly
• [ ] Each major method axis (3 axes in §Methods) closes with a verdict sentence
• [ ] No "neutral catalogue" stretches longer than 3 paragraphs without a verdict / position

---

Equations and Boxes

• [ ] Display equations ($$...$$) appear only in Boxes, not in body paragraphs
• [ ] Textbook formulas (DSC, IoU, FedAvg) handled in prose if not in Box 1, not displayed inline
• [ ] Box 1 (metrics) present and complete
• [ ] Total Box count appropriate for target journal (typically 1-3)

---

Vendor Names

• [ ] No vendor name (HeartFlow, Cleerly, Caristo, Keya, Shukun, etc.) appears in body paragraph
• [ ] All vendor names appear in Table 3 (Commercial Products) only
• [ ] Body uses category descriptors with table cross-reference

# Quick check
for name in HeartFlow Cleerly Caristo "Keya Medical" "Shukun Technology"; do
  grep -n "$name" manuscript_draft.md | grep -v "| $name |"
done
# Expected: 0 hits per name

---

Tables

• [ ] Table 1: Public datasets (year, cases, annotation type, access)
• [ ] Table 2: Method comparison (modality / family / dataset / metric — pick 12-20 papers)
• [ ] Table 3: Commercial products with regulatory evidence
• [ ] Total tables ≤ 4 (typical flagship reviews stay at 2-3)
• [ ] Each table has a title, body, and footnote explaining abbreviations / caveats

---

Figures

• [ ] Figure 1: Review overview / taxonomy
• [ ] Figure 2: Representative architectures or method evolution
• [ ] Figure 3: Clinical workflow or downstream applications
• [ ] (Optional) Figure 4: Data-driven plot (e.g., performance landscape, trial effect sizes)
• [ ] All figures have ≤ 100-word captions following Nature style: bold lead-in title sentence + body sentences
• [ ] No [Figure placeholder] strings before submission

---

Content Coverage

• [ ] All major method axes covered (Architectural / Inductive / Data regime)
• [ ] Negative trials included where they exist (LLM bias: only positive)
• [ ] Inter-vendor reproducibility / cross-site validation discussed where relevant
• [ ] Demographic bias / fairness considerations addressed where relevant
• [ ] Controversies and unresolved questions engaged, not glossed
• [ ] Future directions specific and actionable (not "more research is needed" platitudes)

---

Self-check Commands

# Numbered headings
grep -cE "^#{2,4} [0-9]" manuscript_draft.md
# Expected: 0

# Heading depth violation
grep -c "^#### " manuscript_draft.md
# Expected: 0

# Placeholder DOIs
grep -c "xxx\|x):xxx\|\[TBD\]" manuscript_draft.md
# Expected: 0

# LLM tell phrases
for tell in "has shown promising" "may suggest" "interestingly," "in recent years," "it is worth noting"; do
  echo "=== $tell ==="
  grep -nF "$tell" manuscript_draft.md
done
# Expected: 0 per tell

# Vendor names in body
for vendor in HeartFlow Cleerly Caristo; do
  count=$(grep -c "$vendor" manuscript_draft.md)
  in_table=$(grep "| $vendor " manuscript_draft.md | wc -l)
  echo "$vendor: total $count, in table $in_table, in body $((count - in_table))"
done
# Expected: in body == 0 for each

# Inline equations
grep -n '\$\$' manuscript_draft.md
# All hits should be inside Box context (check 2 lines before)

# Verdict-sentence presence
grep -nE "currently the most|has yet to demonstrate|best understood as|next [0-9]+ years will" manuscript_draft.md
# Expected: ≥ 3 hits

# Citation count (no quantity target — but useful for sanity)
grep -cE "^[0-9]+\." manuscript_draft.md
# Use as input for body↔bib reconciliation

---

What's NOT on this checklist (intentional removals from v2)

• ❌ "Hedging language used" — was actively harmful; hedging-by-default is the LLM tell, not the flagship-review voice
• ❌ "80-120 references" — was driving Claude to pad the bibliography, which encouraged fabrication
• ❌ "Performance metrics consistent (Dice: 0.XXX format)" — checking format ≠ checking correctness
• ❌ "All major methods covered" — was driving exhaustive enumeration over selective synthesis
• ❌ "Recent literature included (>50% from last 3 years)" — date-based filter has no relationship to quality

These were structural illusions of quality. They've been replaced with substantive checks above.

---

Severity Levels for Failures

When a checklist item fails during writing:

Project File Templates

Templates for CLAUDE.md, IMPLEMENTATION_PLAN.md, and the 3 standard tables. v3 updates: Stage 4 (Methods) uses 3-axis structure, not flat 10-subsection list. Bibliography target removed.

---

CLAUDE.md Template

# [Topic] Literature Review Writing Guidelines

## Project info
- **Topic**: [specific topic]
- **Target journals**: [Tier 1: e.g., Nat Rev Cardiol] / [Tier 2: e.g., JACC Cardiovasc Imaging] / [Tier 3: e.g., Eur Radiol]
- **Paradigm spec**: see PARADIGM.md (Phase 0 output)
- **Citation integrity**: see references/CITATION_INTEGRITY.md
- **Hallucination self-check**: see references/HALLUCINATION_PATTERNS.md

## Terminology Standardization

| Unified Term | Avoid Using |
|---|---|
| coronary CT angiography (CCTA) | cardiac CT, CT angio, CTA (ambiguous) |
| centerline (anatomical structure) | midline, vessel line |
| [add domain-specific terms] | [variants to avoid] |

## Canonical Paper / Dataset Descriptions

To prevent internal inconsistency (HALLUCINATION_PATTERNS.md ▸ Pattern 9), lock in canonical descriptions:

| Paper / Dataset | Canonical description |
|---|---|
| ImageCAS | 1000 cases, single-center, CCTA artery segmentation |
| TransCC (Xu et al.) | 3 authors: Chenchu Xu, Meng Li, Xue Wu; Dice 0.730 / IoU 0.582 on ImageCAS |
| FISH&CHIPS | NHS national CT-FFR implementation program; publication: Fairbairn TA et al. Nat Med 2025;31(6):1903-1910 |
| [add high-value papers] | [verified description] |

## Reference Sources

### ArXiv MCP
Search queries used:
- "[topic] segmentation transformer" (cs.CV, eess.IV)
- "[topic] deep learning" (cs.LG)
Date range: last 3 years
Downloaded papers: [list arXiv IDs]

### PubMed MCP
MeSH queries used:
- "Deep Learning"[MeSH] AND "[domain]"[MeSH]
- "[method]"[MeSH] AND "diagnosis"[MeSH]
Filters: Review, Clinical Trial

### Zotero
API: localhost:23119
Relevant collections:
- [Collection name 1]: [N papers]
- [Collection name 2]: [N papers]

## Literature inventory (3-axis grouping)

### Axis 1: Architectural priors
- **CNN-based**: [list verified refs]
- **Transformer-based**: [list]
- **Mamba / state-space**: [list]

### Axis 2: Inductive priors
- **Topology-aware**: [list]
- **Multi-task**: [list]
- **Graph-based**: [list]

### Axis 3: Data regime
- **Self-supervised**: [list]
- **Foundation models**: [list]
- **Federated**: [list]
- **Physics-informed**: [list]

### Clinical / downstream
- [list]

## Verdict positions to take (3-5)

Plan the 3-5 verdict sentences in advance:
- §Methods axis 1 closer: ...
- §Methods axis 2 closer: ...
- §Methods axis 3 closer: ...
- §Clinical translation closer: ...
- (optional) §Discussion closer: ...

## Writing-time guardrails (from references/)

- Per-claim verification on every `[N]` placed (CITATION_INTEGRITY ▸ Rule 1-5)
- Self-scan against 9 hallucination patterns every 5-6 paragraphs (HALLUCINATION_PATTERNS.md)
- Vendor names only in Table 3
- Equations only in Box 1
- Max 2 heading levels, no numbered headings

---

IMPLEMENTATION_PLAN.md Template

# Implementation Plan: [Review Title]

## Phase 0: Paradigm Capture
**Status**: [Not Started / In Progress / Complete]

- [ ] Identify 2-3 exemplar reviews from target tier
- [ ] Read carefully (60-90 min each)
- [ ] Extract style spec to PARADIGM.md
- [ ] Re-anchor at start of each Phase 4 writing session

## Phase 1: Project Initialization
**Status**: [Not Started / In Progress / Complete]

- [ ] CLAUDE.md created with terminology + sources + literature inventory skeleton
- [ ] IMPLEMENTATION_PLAN.md (this file) created
- [ ] manuscript_draft.md created (empty)
- [ ] PARADIGM.md linked

## Phase 2: Literature Collection + Verification
**Status**: [Not Started / In Progress / Complete]

### ArXiv (Deep Learning Methods)
- [ ] Search "[topic] segmentation" in cs.CV, eess.IV
- [ ] Search "[topic] transformer/attention" in cs.CV
- [ ] Download key papers (target: 50-80 — discriminate aggressively)
- [ ] Read each (abstract + methods + results)
- [ ] Note actual module names + actual numbers in working notes

### PubMed (Clinical Literature)
- [ ] Search MeSH: "Deep Learning" AND "[domain]"
- [ ] Filter by publication type (Review, Clinical Trial)
- [ ] Collect clinical validation studies (target: 20-40)
- [ ] PubMed WebFetch on each for metadata

### Zotero (User's local library)
- [ ] Check existing collections for closed-access papers
- [ ] Note Zotero item keys for later fulltext retrieval

### Additional sources
- [ ] WebSearch supplementary (IEEE Xplore, Springer)

### Verification (every entry, before adding)
- [ ] DOI resolves on Crossref
- [ ] First + last author verbatim from first-source
- [ ] Journal, year, vol, issue, pages match
- [ ] No `xxx` / `[TBD]` placeholders

### Gap analysis
- [ ] Negative trials covered? (LLM bias: only positive)
- [ ] Inter-vendor reproducibility covered?
- [ ] Demographic-bias studies covered?
- [ ] Recent 6-month preprints covered?

## Phase 3: Outline + 3-Axis Taxonomy
**Status**: [Not Started / In Progress / Complete]

- [ ] Top-level sections defined (Introduction / Methods / Applications / Translation / Discussion)
- [ ] §Methods structured as **3 H3 subsections** (3 axes):
  - [ ] §Methods.Architectural priors
  - [ ] §Methods.Inductive priors
  - [ ] §Methods.Data regime
- [ ] Each paper in literature inventory mapped to an axis
- [ ] Table 1 / 2 / 3 designs planned
- [ ] Box 1 (metrics) content planned
- [ ] Figures (3-5) planned
- [ ] Verdict positions for 3 axes selected

## Phase 4: Write with Per-Claim Verification
**Status**: [Not Started / In Progress / Complete]

For each section, micro-workflow (per CITATION_INTEGRITY.md):
- [ ] Re-read the actual cited papers
- [ ] Write 2-4 sentences with actual module names + actual numbers
- [ ] Verify each `[N]` placed (body↔bib, number, direction)
- [ ] Close axis with verdict sentence
- [ ] Equations to Box 1, not body
- [ ] Vendor names to Table 3, not body
- [ ] Self-scan against 9 hallucination patterns every 5-6 paragraphs

### Section progress

- [ ] Introduction (Clinical background / Technical challenge / Scope)
- [ ] Datasets and evaluation metrics (Box 1, Table 1)
- [ ] Methods §Axis 1: Architectural priors
- [ ] Methods §Axis 2: Inductive priors
- [ ] Methods §Axis 3: Data regime
- [ ] Methods Table 2 finalized
- [ ] Downstream applications
- [ ] Translation to clinical practice (Table 3)
- [ ] Outstanding challenges
- [ ] Future directions
- [ ] Conclusion
- [ ] References (cross-checked with body)

## Phase 5: Multi-Agent Peer Review
**Status**: [Not Started / In Progress / Complete]

- [ ] TeamCreate `manuscript-review`
- [ ] Spawn `style-reviewer` teammate (Task 1)
- [ ] Spawn `ref-checker` teammate (Task 2)
- [ ] Spawn `peer-reviewer` teammate (Task 3)
- [ ] Spawn `fact-checker` teammate (Task 4)
- [ ] Synthesize 4 reports into 00_team_synthesis.md
- [ ] Address all CRITICAL findings
- [ ] Address HIGH findings agreed by ≥ 2 reviewers

## Phase 6: Submission Prep
**Status**: [Not Started / In Progress / Complete]

- [ ] Journal selected (reach + match + safety tiers)
- [ ] Presubmission inquiries sent (reach tier, optional)
- [ ] Cover letter drafted
- [ ] Box vs body duplication scan run
- [ ] Section cross-reference scan run
- [ ] Figures produced (not placeholders)
- [ ] Citation format converted for target journal
- [ ] Final self-check checklist passed

## Literature Sources Summary

| Source | Query / Collection | Papers | Verified | Status |
|---|---|---|---|---|
| ArXiv | [query 1] | N | N | [ ] |
| ArXiv | [query 2] | N | N | [ ] |
| PubMed | [MeSH query] | N | N | [ ] |
| Zotero | [collection] | N | N | [ ] |

## Change log

### [Date] - v1.0
- Initial draft completed through Phase 4
### [Date] - v1.1
- Phase 5 peer review found 8 issues; addressed
### [Date] - v1.2
- Phase 6 submission prep done; sent to [Journal]

---

Standard Table Templates

Table 1: Public Datasets

**Table 1 | Public datasets for [task].**

| Dataset | Year | Cases | Annotation type | Access |
|---|---|---|---|---|
| [Name] | 20XX | N | [type] | [link] |

*Cases = unique patients. Where multi-center, the number of sites is given in parentheses. Access notations: O = open, R = restricted (registration), C = commercial.*

Table 2: Method Comparison

**Table 2 | Representative deep learning methods for [task].**

| Reference | Modality | Family | Architecture | Dataset | Headline metric | Innovation |
|---|---|---|---|---|---|---|
| [First-author Year] [N] | CCTA | Topology-aware | nnU-Net + clDice | ImageCAS | Dice 0.812 | First clDice on CCTA |

*Performance numbers are taken from the original publications and should not be compared directly across rows; datasets, splits and metric definitions differ. Family is the most-engaged inductive prior or training regime, not the only one. CCTA = coronary CT angiography; HD95 = 95th-percentile Hausdorff distance; — = not reported.*

Table 3: Commercial Products

**Table 3 | Commercial AI tools for [domain] with regulatory clearance.**

| Manufacturer | Product | Indication | Underlying technology | Regulatory | Deployment | Key validation evidence |
|---|---|---|---|---|---|---|
| HeartFlow | CT-FFR Analysis | CT-FFR | CFD + DL hybrid | FDA cleared (CPT 75580); CE | Cloud-based SaaS | NXT, ADVANCE, PLATFORM, PACIFIC; FISH&CHIPS [N] |
| Keya Medical | DeepVessel FFR | CT-FFR | DL surrogate of CFD | FDA, NMPA Class III, CE | Cloud + on-site | Multi-centre Chinese validation [N] |

*Listed validation evidence refers to peer-reviewed primary publications, not vendor white papers. NMPA = National Medical Products Administration (China); CFD = computational fluid dynamics.*

---

Notes

Why bibliography target removed

v2 set "80-120 references" as a quality criterion. This drove Claude to pad the bibliography, and padding encouraged fabrication. v3 has no count target. Cite what supports the argument. If your review naturally has 60 high-quality refs, that's enough. If it needs 150, also fine. Quantity is downstream of substance.

Why 3-axis Methods is the structural commitment

v2's coronary review produced 10 flat method H3 subsections. The resulting §Methods read as a textbook chapter, not a flagship review. Flagship reviews compress 10+ method variants into ~3 thematic axes that force comparative synthesis. The 3-axis structure is the most important structural fix in v3.

6-Phase Literature Review Workflow

This workflow replaces the v2 "7-phase write-then-QA" structure with a "write-with-verify + multi-agent peer review" structure. Verification is embedded throughout, not bolted on at the end.

---

Phase 0: Paradigm Capture

Goal: Before writing anything, ground the review in 2-3 published flagship-quality exemplars from the target journal tier.

Why: LLMs default to a generic "survey paper" register — numbered chapters, hedging language, neutral catalogue, dense subsections. Flagship reviews (Nature Reviews / Nat Med / Lancet family) write very differently. Without an exemplar to anchor against, the draft will drift toward survey-paper style and become hard to retrofit.

Actions:

1. Identify 2-3 exemplar reviews from the target journal tier. Selection criterion: same modality or same problem family, published in last 3 years, by recognized authors.

2. Read them carefully (not just skim). Pay attention to:

• Heading depth and numbering (most flagship reviews use 2 levels, no numbers)
• Paragraph rhythm (topic → evidence → verdict, not topic → list → list → list)
• Citation density (1.5-2.5 refs per paragraph typical)
• Equation handling (Box vs inline)
• Vendor handling (Table vs body)
• Authorial voice (do they take positions or stay neutral?)
• Box and Figure usage (how many, how dense)
• Table count (rarely more than 3)

3. Write the extracted style spec to PARADIGM.md. See PARADIGM.md template for the exact structure.

Deliverable: PARADIGM.md in the project root.

Time budget: 2-3 hours (reading + spec writing).

---

Phase 1: Project Initialization

Goal: Set up the project files and writing guidelines.

Actions:

1. Create CLAUDE.md from the TEMPLATES.md ▸ CLAUDE.md template. Fill in:

• Topic-specific terminology table
• Reference source configurations (ArXiv / PubMed / Zotero queries)
• Expected literature categories (don't commit yet — refine after Phase 2)

2. Create IMPLEMENTATION_PLAN.md from TEMPLATES.md ▸ IMPLEMENTATION_PLAN.md. Note the 3-axis structure for §Methods (not flat).

3. Create empty manuscript_draft.md. Leave it empty until Phase 4 — don't pre-populate with placeholder text.

4. Link PARADIGM.md and CITATION_INTEGRITY.md and HALLUCINATION_PATTERNS.md from CLAUDE.md so they're easy to refer back to.

Deliverable: Project skeleton with 4 files (PARADIGM.md, CLAUDE.md, IMPLEMENTATION_PLAN.md, manuscript_draft.md).

Time budget: 1 hour.

---

Phase 2: Literature Collection + Verification (simultaneous)

Goal: Gather the corpus while verifying each entry's metadata in real time.

Why simultaneous: v2 separated collection from verification — collection in Phase 2, verification never. Result: 17 placeholder DOIs and many wrong-author lists shipped to the final draft. v3 verifies on the way in.

Actions:

2.1 ArXiv (methodological preprints)

Query: "[topic] AND (segmentation OR detection OR classification)"
Categories: cs.CV, eess.IV, cs.LG
Date: last 3 years
Max results: 50-80 per query (NOT 100 — discriminate aggressively)

For each paper added:

• mcp__arxiv-mcp-server__download_paper(paper_id)
• mcp__arxiv-mcp-server__read_paper(paper_id) — read abstract + methods + results
• Note: actual title, full author list, exact module names, headline numbers, dataset name

2.2 PubMed (peer-reviewed clinical literature)

MeSH: "Deep Learning"[MeSH] AND "[domain]"[MeSH]
Filters: Review or Clinical Study, last 5 years

For each paper added:

• WebFetch on https://pubmed.ncbi.nlm.nih.gov/<PMID>/
• Note: title, full author list (first 6 + et al.), journal, year, volume, issue, pages, DOI, headline finding direction (positive/negative)

2.3 Zotero (user's local library — especially closed-access)

For closed-access journals (Med Image Anal, Eur Radiol, JACC, Lancet family, Nature family) the user often has PDFs in Zotero. Always check before assuming inaccessible.

mcp__zotero__zotero_search_collections(query: "<topic>")
mcp__zotero__zotero_get_collection_items(collection_key: ..., detail: "summary", limit: 200)
mcp__zotero__zotero_search_items(query: "<author> <method>", limit: 5)
mcp__zotero__zotero_get_item_fulltext(item_key: ...)

2.4 Verification — apply CITATION_INTEGRITY rules as you collect

For each paper added to the bibliography, before committing the entry:

• [ ] DOI resolves on Crossref (api.crossref.org/works/<DOI>)
• [ ] First and last author names match the first-source verbatim
• [ ] Journal name, year, volume, issue, pages match
• [ ] No placeholder strings (xxx, [TBD], ?) anywhere in the entry

If any check fails, do not add the entry. Either resolve the metadata or drop the paper.

See CITATION_INTEGRITY.md for the full 5-rule protocol.

2.5 Build the literature matrix

Note the 3 axes, not a flat 10-category list. See DOMAINS.md for the 3-axis structure per domain.

2.6 Gap analysis

After initial collection:

• Are negative trials covered? (LLM tendency: only cite positive trials → over-rosy review)
• Are recent 6-month preprints covered? (LLM tendency: stale by training-cutoff date)
• Are inter-vendor reproducibility studies covered? (LLM tendency: report-only-positives)
• Are demographic-bias / fairness studies covered? (LLM tendency: ignore entirely)

For each gap, run a targeted search.

Deliverable: Literature matrix (in CLAUDE.md or IMPLEMENTATION_PLAN.md) with every entry verified.

Time budget: 1-2 days (depending on topic breadth). Most of the time is reading abstracts to discriminate relevance, not searching.

---

Phase 3: Outline + 3-Axis Method Taxonomy

Goal: Lock in section structure + the 3-axis method grouping before writing prose.

Actions:

1. Define top-level sections from the Standard Review Structure template. No numbered headings.

2. For §Methods, force yourself into the 3-axis grouping (NOT a flat 10-subsection list):

• Axis 1: Architectural priors (what kind of network)
• Axis 2: Inductive priors (what kind of geometric / structural / multi-task bias is built in)
• Axis 3: Data regime (how is data used / pre-trained / federated)

Each axis becomes one H3 subsection. Inside each axis, group method families with bold lead-ins (**Topology-aware design.**), not deeper H4 headings.

3. Map each paper from the literature matrix to one (or sometimes two) axes. A paper that's only about a topology loss → axis 2. A paper that's a Mamba variant for segmentation → axis 1. A paper about federated learning for cardiac MRI → axis 3.

4. Plan the three tables explicitly:

• Table 1: public datasets (name, year, cases, annotation type, access)
• Table 2: representative methods (with modality / family / dataset / metric — pick 12-20 papers across all 3 axes)
• Table 3: commercial products (manufacturer / product / regulatory / validation evidence)

5. Plan Box 1: evaluation metrics with formulas.

6. Plan figures (typically 3-5; overview, taxonomy, representative architectures, workflow).

Deliverable: Section outline and 3-axis paper mapping in IMPLEMENTATION_PLAN.md.

Time budget: 0.5-1 day.

---

Phase 4: Write with Per-Claim Verification

Goal: Produce the manuscript prose, with verification embedded in every paragraph.

Why per-claim verification: v2 wrote first and QA'd later. The QA was structural, not factual. Result: shipped fabricated module names and wrong numbers. v3 verifies each claim before committing it.

Actions per section:

1. Write an introduction paragraph (1-2 paragraphs on motivation + scope). This is the safest part of the section — make it punchy and clear, set up the verdict that will close the section.

2. For each method family, repeat this micro-loop:

a. Re-read the cited paper's abstract + relevant methods/results section. Use read_paper (arxiv) or zotero_get_item_fulltext (closed-access). Do not skip this. If you can't access the paper, do not write its internal architecture.

b. Write 2-4 sentences describing the method's actual contribution, using actual module names and actual numbers. Cite the paper as [N].

c. Verify the [N] you just placed:

• Is N's bibliography entry the paper you just read? (body↔bib reconciliation)
• Does your sentence's number (Dice / sensitivity / HR) appear in the paper?
• Does your sentence's directional claim (lower/higher / increased/decreased) match the paper?

d. If any verification fails, fix immediately. Do not move on with broken citations — they compound.

3. Close the section with a verdict sentence (for 3 of the H3 subsections — pick the most opinionated positions). See SKILL.md ▸ Verdict Sentences.

4. Equations go into Box 1, not the body. If you find yourself typing a $$ outside Box 1, stop and move the equation.

5. Vendor names go into Table 3, not the body. If you find yourself typing "HeartFlow" / "Cleerly" / etc. outside Table 3, stop and rewrite with category descriptor.

6. Update bibliography as you go (don't batch at the end — the body↔bib reconciliation breaks down with batching).

Self-check during writing

Every 5-6 paragraphs, pause and scan for the 9 hallucination patterns (see HALLUCINATION_PATTERNS.md):

• Are author names sounding generic? (pattern 1)
• Are performance numbers suspiciously round or high? (pattern 2)
• Am I claiming directional findings I haven't verified? (pattern 3)
• Am I citing vendor materials as if peer-reviewed? (pattern 4)
• Are there any xxx or [TBD] strings? (pattern 5)
• Are [N] referring to what I think? (pattern 7)
• Are my metric formulas correct? (pattern 8)

Deliverable: manuscript_draft.md complete from Introduction through References, with per-claim verification trace in the writing log.

Time budget: 3-5 days (the largest phase).

---

Phase 5: Multi-Agent Peer Review

Goal: Before delivering to the user, run a 4-perspective audit.

Why: Single-author self-review misses patterns. The 4 specialized agents catch what a single writer doesn't.

Actions:

Launch a manuscript-review agent team with 4 teammates (see the ai-review-revision skill's references/agent_team_setup.md for the exact TeamCreate / TaskCreate / Agent spawn templates):

After all 4 reports return, synthesize into review_outputs/00_team_synthesis.md. For any issue ≥2 reviewers independently flag, treat as hard fix. For single-reviewer flags, judge based on severity.

If the issues are minor (handful of style nits, 1-2 minor citation issues): fix them in place and ship.

If the issues are major (≥5 hard factual errors, ≥10 citation drift instances, missing major controversies): the right move is to apply the ai-review-revision skill's Phase 1 (factual reset) workflow before delivery. This is unusual but possible — and it's a signal that this skill needs further improvement.

Deliverable: 4 review reports + synthesis + a final-quality draft.

Time budget: 1 day for the multi-agent review, plus fix time depending on findings.

---

Phase 6: Submission Prep

Goal: Format-level finalization for a specific target journal.

Use the ai-review-revision skill's references/phase4_submission_prep.md for:

• Journal selection (3-tier reach / match / safety)
• Presubmission inquiry templates
• Cover letter template
• Box vs body duplication scan
• Figure realization (no placeholders at submission)
• Citation format conversion ([N] → <sup>N</sup> depending on journal)
• Self-check checklist

Deliverable: Submission-ready package (manuscript + figures + cover letter + author info).

Time budget: 0.5-1 week including presubmission inquiry wait.

---

Total Timeline

For a typical medical-imaging review project:

Compare to v2 + downstream fix: typically 1 week of v2 drafting + 2-3 weeks of post-hoc revision. Net the same time, but v3 delivers a submission-ready draft instead of one needing factual reset.