architecture-zoo · diff

git:20260725.58283ba to git:20260725.12755d7

1 added, 1 removed. Audit A to A.

---

name: architecture-zoo

description: >

Choose a model architecture for a medical-imaging research question before scaffolding. Maps the task

(classification, segmentation, detection, transfer), modality and dimensionality, labelled-data scale,

and class imbalance to a shortlist of architectures, each grounded in its source paper with a

when-to-use, a medical-imaging use, a reference implementation, the typical validation setup, and the

matching model-scaffold template. Covers the foundational curriculum (ResNet, DenseNet, EfficientNet,

ViT, Swin; U-Net, 3-D U-Net, Attention/Residual U-Net, nnU-Net (+ ResEnc/MedNeXt/STU-Net), Mask R-CNN;

SAM/MedSAM, nnInteractive/VISTA3D interactive-3D, TotalSegmentator, BiomedCLIP, DINO/MAE/SimCLR; and

graph neural nets — GCN/GraphSAGE/GAT/GIN/BrainGNN — for brain connectomes). It teaches archetypes and

the task-to-architecture logic (including the "scale the CNN, new≠better" rigour caveat), not a live

SOTA leaderboard.

- triggers: architecture zoo, which architecture, choose a model, model selection, ResNet vs ViT, U-Net vs nnU-Net, what backbone, foundation model for, transfer learning choice, MedSAM, TotalSegmentator, DINO, MAE, self-supervised, graph neural network, GNN, brain connectome, GCN, GAT, GraphSAGE, BrainGNN, population graph, paper to architecture, reference implementation, when to use ViT, segmentation architecture, classification backbone, nnU-Net ResEnc, MedNeXt, STU-Net, nnInteractive, VISTA3D, SAM-Med3D, Mamba, U-Mamba, interactive segmentation, labelling acceleration, promptable segmentation

+ triggers: architecture zoo, which architecture, choose a model, model selection, ResNet vs ViT, U-Net vs nnU-Net, what backbone, foundation model for, transfer learning choice, MedSAM, TotalSegmentator, DINO, MAE, self-supervised, graph neural network, GNN, brain connectome, GCN, GAT, GraphSAGE, BrainGNN, population graph, paper to architecture, reference implementation, when to use ViT, segmentation architecture, classification backbone, nnU-Net ResEnc, MedNeXt, STU-Net, nnInteractive, VISTA3D, SAM-Med3D, Mamba, U-Mamba, interactive segmentation, labelling acceleration, promptable segmentation, nnDetection, lesion detection, ConvNeXt, YOLO, YOLOv8, RT-DETR, DETR, RetinaNet, detection architecture

tools: Read, Write, Edit, Grep, Glob

model: inherit

---

# Architecture-Zoo Skill

## Purpose

This skill turns a **medical-imaging research question into a paper-grounded architecture choice** —

so the build starts from the right archetype (and a known validation setup) rather than from whatever is

fashionable, and the choice carries its source citation into the Methods. It is the **front end** of the

model-engineering lane: `architecture-zoo (choose)` → `/model-scaffold (build)` → `/model-validation

(validate)`.

It is **advisory** (Layer D): it writes a short decision note, never code or weights. The actual repo is

`/model-scaffold`. It describes **archetypes and the task → family → constraint logic**, not a live SOTA

leaderboard (SOTA churns; the logic does not).

## When to use

- You need to pick an architecture/backbone for a classification, segmentation, detection, or

transfer-learning question and want it grounded in the literature with a sensible default.

## When NOT to use

- Generating the runnable repo → `/model-scaffold`.

- Auditing a trained model's validation design → `/model-validation`.

- Metrics / calibration → `/model-evaluation` + `/analyze-stats`.

- General study/validity design → `/design-study`; AI-vs-expert benchmark → `/design-ai-benchmarking`.

- LLM / MLLM → `/mllm-eval`.

## Workflow

### Phase 1 — Frame the question

State the **task** (classification / segmentation / detection / transfer), the **modality +

dimensionality** (2-D vs 3-D volume), the **labelled-data scale** (events / structures, not just

images), **label availability** (lots / few / unlabelled pool), and constraints (class imbalance,

small structures, interpretability, deployment compute).

### Phase 2 — Walk the decision tree

Open `${CLAUDE_SKILL_DIR}/references/index.md` and follow task → constraints → default pick. It routes to

a family card.

### Phase 3 — Read the family card

- `${CLAUDE_SKILL_DIR}/references/classification.md` — ResNet / DenseNet / EfficientNet / Inception /

ViT / Swin / DeiT.

- `${CLAUDE_SKILL_DIR}/references/segmentation.md` — U-Net / 3-D U-Net / V-Net / Attention & Residual

U-Net / nnU-Net / SegResNet / Swin-UNETR / Mask R-CNN.

- `${CLAUDE_SKILL_DIR}/references/detection.md` — R-CNN family / Faster R-CNN + FPN / Mask R-CNN /

RetinaNet / YOLO / DETR.

- `${CLAUDE_SKILL_DIR}/references/synthesis.md` — Pix2Pix / CycleGAN / SPADE / diffusion (DDPM, latent) /

VAE / fastMRI reconstruction.

- `${CLAUDE_SKILL_DIR}/references/foundation_models.md` — SAM / MedSAM / MedSAM2 / TotalSegmentator /

SegVol / BiomedCLIP / DINO / MAE / SimCLR / MoCo.

- `${CLAUDE_SKILL_DIR}/references/graph.md` — GCN / GraphSAGE / GAT / GIN / BrainGNN for brain

connectomes & population graphs (integrate PyTorch Geometric / DGL; not scaffolded by model-scaffold).

Each card gives the paper, core idea, when-to-use, medical-imaging use, reference implementation, and the

**typical validation/experiment setup** for that architecture class.

### Phase 4 — Write the decision note

Record `decisions/architecture_choice.md`: the **task**, the **chosen architecture**, its **source

paper**, the **reason** against the constraints, the **runner-up + why not**, and the matching

**`/model-scaffold` template**. Naming the source paper is mandatory; cite, never invent, any benchmark

number.

### Phase 5 — Hand off

Carry the decision note to `/model-scaffold` (instantiate the template), then `/model-validation`

(split / validation design), `/model-evaluation` + `/analyze-stats` (metrics), and `/write-paper`

(the Methods cite the architecture's source paper).

## Anti-Hallucination

- **Never recommend an architecture without naming its source paper.** Every card cites the paper; the

decision note must carry that citation.

- **Never invent benchmark numbers or paper claims.** If a number matters, cite it (verify via

`/search-lit`); if uncertain, write `[VERIFY]` and ask.

- **Never recommend an architecture for a modality or data scale it does not suit** (e.g. a from-scratch

ViT on a few hundred images, or 2-D slices for a volumetric structure) — the constraints in the

decision tree exist to prevent exactly that.

- The zoo is a curated **archetype** map, not a current SOTA ranking — say so rather than implying a

recommendation is the latest best.

## Boundaries

```

architecture-zoo (this skill: choose, paper-grounded)

└─ model-scaffold (build the reproducible repo from the chosen template)

└─ model-validation -> model-evaluation -> write-paper (cite the source paper)

```

It does not build, train, evaluate, or rank live SOTA — it maps the research question to a defensible,

paper-grounded archetype and hands the choice to `/model-scaffold`.