[ English | 中文 ]

Large-scale, fully autonomous medical image segmentation research — dataset in, paper out.

---

📄 Paper

Camyla: Scaling Autonomous Research in Medical Image Segmentation

Yifan Gao¹², Haoyue Li¹, Feng Yuan¹, Xin Gao¹*, Weiran Huang²³*, Xiaosong Wang²⁴*

^{¹ USTC · ² Shanghai Innovation Institute · ³ SJTU · ⁴ Shanghai AI Lab · * Corresponding authors}

📑 Read the paper (PDF) · alphaXiv · arXiv

Headline results (28 days, zero human intervention):

• 40 complete manuscripts written end-to-end
• Cost-efficient: only $20–30 per paper in LLM API spend
• Beats the strongest per-dataset baseline (chosen from 14 established architectures including nnU-Net) on 24 of 31 datasets under identical training budgets
• CamylaBench: contamination-free benchmark of 31 datasets, built exclusively from 2025 publications
• Stronger long-horizon orchestration on the experiment stage. During experiment execution, driven by cost-efficient backends (GLM-4.7 + MiniMax-M2.5), Camyla outperforms AI Scientist, autoresearch Claude Code (Opus 4.6), and autoresearch Codex (GPT-5.4-xhigh) on execution success, completion rate, and fidelity to the original proposal

Manuscript quality — double-blind evaluation. Camyla-generated manuscripts were mixed with real 2025 publications and judged without reviewers knowing which was AI-written. Four independent panels — 5 senior reviewers, 10 junior reviewers, 5 different AI models, and the Stanford Agentic Reviewer — all place Camyla's output between the T1 and T2 tier of contemporary medical-imaging journals. The T1 anchor is IEEE TMI / Medical Image Analysis; the T2 band is JCR Q1 medical-imaging journals:

Tier	Journals	Papers	Representative venues
T1 (top-tier)	2	10	IEEE Transactions on Medical Imaging; Medical Image Analysis
T2 (JCR Q1)	7	35	IEEE Journal of Biomedical and Health Informatics; Artificial Intelligence in Medicine; et al.
T3	9	45	International Journal of Computer Assisted Radiology and Surgery; Biomedical Physics & Engineering Express; et al.
Total	18	90

---

📰 News

• 2026-04-13 — Initial public release of the code, the paper, and CamylaBench (31 pre-formatted datasets, 📦 Google Drive). We are still cleaning the full CamylaTrace-232K trajectory dataset (per-run logs + intermediate artifacts); release will follow shortly.

---

Camyla is an automated research pipeline that takes a medical image segmentation task, searches the literature for relevant ideas, proposes research hypotheses, runs end-to-end deep-learning experiments (Baseline → Creative Research → Ablation), and writes up the results as a publication-ready paper.

It combines:

• Quality-Weighted Branch Exploration (QWBE) over experiment configurations
• OpenHands-driven code generation and iterative debugging
• Multi-source literature search (ArXiv, OpenAlex, PubMed, Semantic Scholar)
• A paper agent that drafts, compiles, and cites LaTeX papers (Elsevier format supported)
• A flexible LLM routing layer (llm_endpoints + llm_roles) so you can mix providers (OpenRouter, GLM, MiniMax, …) without touching code

Status. Research prototype, preparing for open-source release. APIs may change.

---

📚 Sample generated papers

A 10-paper subset of the manuscripts Camyla produced end-to-end — no hand-editing, LaTeX compiled as-is by the pipeline. Each PDF lives under assets/paper_pdf/.

#	Title	Modality / task
1	Cross-Directional Feature Lattice for Brain Tumor Segmentation	MRI · brain tumor
2	Scale-Frequency Adaptive Fusion for Multiple Sclerosis Lesion Segmentation	MRI · MS lesions
3	Hierarchical Context Gating for Neonatal Brain Lesion Segmentation	MRI · neonatal HIE
4	Cross-Scale Mutual Refinement for Bronchoalveolar Lavage Fluid Cell Segmentation	Microscopy · BALF cells
5	Symmetry-Aware Cascaded Attention for Panoramic Tooth Segmentation	Dental X-ray · tooth
6	Specular-Residual Decoupled Encoding for Surgical Scene Segmentation	Laparoscopy · surgical scene
7	Adaptive Scale-Aware Feature Integration for Liver Lesion Segmentation	CT · liver lesion
8	Boundary-Hierarchical Decomposition for Fetal Brain Tissue Segmentation	MRI · fetal brain
9	Vessel-Guided Boundary Residual Networks for Dermatological Vessel Segmentation	OCTA · dermatological vessel
10	Hierarchical Resolution-Retentive Feature Encoding for Brain Metastasis Segmentation	MRI · brain metastasis

---

Quick start

1. Install

git clone https://github.com/yifangao112/Camyla.git camyla
cd camyla
python -m venv .venv && source .venv/bin/activate   # or: conda create -n camyla python=3.10
pip install -r requirements.txt

# Sister packages (segmentation framework + raw-data conversion agent)
pip install git+https://github.com/yifangao112/CamylaNet.git
pip install git+https://github.com/yifangao112/nnPrep.git

A Python 3.10+ environment is expected. GPU is required for actually running the segmentation experiments (baseline uses CamylaNet / nnU-Net v2).

2. Configure

Set CamylaNet's data-path environment variables (same convention as CamylaNet's own README — the baseline pipeline reads and writes under these paths):

export camylanet_raw="/path/to/camylanet_raw"
export camylanet_preprocessed="/path/to/camylanet_preprocessed"
export camylanet_results="/path/to/camylanet_results"   # baseline artifacts land here

Then copy the example Camyla config and edit it:

cp config_example.yaml config.yaml

At minimum, set:

• At least one entry under llm_endpoints with a valid api_key (or export the matching api_key_env, e.g. OPENROUTER_API_KEY)
• default_endpoint — the endpoint name most roles will use

See Configuration below for the full layout, role-routing semantics, and N-way model competition. The shipped config_example.yaml already has a complete multi-provider setup you can copy from.

3. Prepare the dataset

Camyla operates on a dataset that already lives in nnU-Net v2 layout under $camylanet_raw/Dataset{ID}_{Abbr}/. You have three options:

• Use CamylaBench (recommended). The 31 datasets used in the paper are pre-formatted and ready to drop in — download from 📦 Google Drive and extract each Dataset{ID}_{Abbr}/ folder under $camylanet_raw/. The ready-made idea descriptions in ideas/ match these IDs.
• Bring your own, already in nnU-Net v2 format. Just point camylanet_raw at it.
• Bring your own raw data in some arbitrary layout. Use nnPrep (an LLM agent that converts arbitrary medical segmentation datasets into the nnU-Net v2 format) — or write your own conversion script.

4. Run

Pick an idea from ideas/ (31 ready-to-use dataset descriptions) and launch:

python launch_camyla.py \
    --config config.yaml \
    --load_ideas ideas/900.json \
    --idea_idx 0

On first run for a given dataset, Camyla automatically runs the baseline pipeline (trainer screening + full training for everything that passes) under $camylanet_results. Subsequent runs reuse the artifacts and skip this step.

The pipeline writes everything to experiments/<date>_<idea>_attempt_<id>/:

experiments/2026-04-12_liver_segmentation_attempt_0/
├── idea.json / idea.md           # task spec
├── config.yaml                   # resolved config used for this run
├── logs/0-run/
│   ├── experiment_report.md      # human-readable summary
│   └── experiment_results/       # metrics, checkpoints, plots
├── research_proposals/           # auto-generated proposals
└── paper/                        # LaTeX + compiled PDF (if writeup enabled)

Common flags:

Flag	Purpose
--resume_from_checkpoint PATH	Resume from a previous checkpoint.pkl
--skip_writeup / --skip_review	Run experiments only, skip paper generation
--debug-baseline	Fake the baseline metrics so Stage 2 runs immediately (dev only)
--verbose	DEBUG-level logging

---

How it works

• Phase 1-3 (idea generation). Searches 1-4 literature sources, extracts open research challenges, and generates multiple proposals scored by an assessment LLM.
• Stage 1-3 (experiment). QWBE expands a tree of code variants. Stage 2 can run N-way model competition by listing multiple experiment.code.candidates.
• Paper Agent. Takes the final experiment results and produces a cited paper.

---

Configuration

Every LLM call Camyla makes resolves through a two-layer system:

1. llm_endpoints — your named LLM connections (one per provider/backend).
2. llm_roles — every internal component ("role") picks an endpoint and optionally overrides its model, temperature, or max_tokens.

Anything you don't configure at the role level falls back to default_endpoint.

1. llm_endpoints — connections

Each entry is an OpenAI-compatible endpoint. The shape is fixed:

llm_endpoints:
  my_openrouter:
    api_key: ""                              # inline key (leave empty → use env)
    api_key_env: OPENROUTER_API_KEY          # env var name to read when api_key is empty
    base_url: "https://openrouter.ai/api/v1"
    model: "deepseek/deepseek-v3.2"          # default model for this endpoint
    temperature: 0.5                         # default temperature

Key resolution order: non-empty api_key > environment variable named by api_key_env > empty (error at first call).

Naming is free-form. my_openrouter, cheap_backend, gpt4 — all valid. Roles reference endpoints by name, so changing models across the whole pipeline is just swapping one string. Common backends we've tested:

Backend	base_url	Notes
OpenRouter	https://openrouter.ai/api/v1	one key, 300+ models
DashScope (Qwen)	https://coding.dashscope.aliyuncs.com/v1	cheap Qwen/GLM routing
MiniMax	https://api.minimaxi.com/v1	M-series models
OpenAI	https://api.openai.com/v1	native
Local vLLM / Ollama	http://localhost:8000/v1	any OpenAI-compatible server

2. default_endpoint

The global fallback. Every role without an explicit endpoint uses this one.

default_endpoint: my_openrouter

3. llm_roles — per-role overrides

A role is one logical LLM use inside the pipeline (feedback, paper writer, idea generator, etc.). You only specify the fields you want to override:

llm_roles:
  # Tree-search roles
  feedback:           { temperature: 0.9, max_tokens: 8192 }
  log_summary:        { temperature: 1.0 }

  # Idea-generation roles — swap to a cheaper model for these
  literature_backbone:  { model: google/gemini-3-flash-preview }
  challenge_extraction: { temperature: 0.3 }

  # Paper agent sub-agents: `_default` applies to the whole group unless
  # an individual sub-agent overrides it.
  paper_agent:
    _default:             { temperature: 0.6 }
    BibtexAgent:          { model: z-ai/glm-4.7, temperature: 0.3 }
    IdeaGenerationAgent:  { model: google/gemini-3-flash-preview, temperature: 0.8 }

  # Paper writing roles can be routed to a different endpoint entirely.
  paper_writing:
    latex_editor:    { endpoint: my_dashscope, temperature: 0.7 }
    image_generator: { endpoint: my_openrouter,
                       model: google/gemini-3.1-flash-image-preview,
                       aspect_ratio: "16:9", image_size: "2K" }

Override precedence for any given role: role fields > its endpoint's defaults. You can point a role at a completely different endpoint with endpoint: <name>.

4. N-way model competition (Stage 2)

Under experiment.code.candidates, list the endpoint names that should compete as code authors for Stage 2. Camyla will run one branch per candidate and keep the strongest.

experiment:
  code:
    candidates: [my_dashscope, my_minimax]
    max_tokens: 16384

Set it to a single-element list to disable competition.

5. Non-LLM API keys

Literature search keys live under api_keys (same inline/env fallback pattern):

api_keys:
  s2:   { value: "", env: S2_API_KEY }    # Semantic Scholar
  ncbi: { value: "", env: NCBI_API_KEY }  # PubMed

Camyla works without these, but rate limits will be tighter.

6. Tuning the run

The rest of the config controls what Camyla does, not which LLMs it uses:

• idea_generation.* — how many papers to search, how to score proposals, how many generator personalities to ensemble
• experiment.stages.* — per-stage iteration budgets (Stage 1 = baseline replication, Stage 2 = creative research, Stage 3 = ablation)
• experiment.openhands.* — OpenHands coder settings (python path, iteration cap, condenser)
• experiment.search.* — QWBE hyperparameters (UCB constant, debug probability, draft count)

All of these have sensible defaults in config_example.yaml; you usually only touch stages.*_max_iters when you want a faster / cheaper run.

---

Repository layout

camyla/
├── LICENSE
├── README.md                        # this file
├── config_example.yaml              # documented config template
├── requirements.txt
├── launch_camyla.py                 # main entry point
├── ideas/                           # 31 ready-made idea JSONs
└── camyla/                          # core package
    ├── model_config.py              # LLM config loader (get_endpoint/get_role)
    ├── baseline/                    # screening + full training before QWBE
    ├── infrastructure/literature/   # arxiv / openalex / pubmed / multi-source
    ├── paper_agent/                 # LaTeX writer + plotters + bibtex agent
    ├── tools/                       # OpenAlex / Semantic Scholar tools
    ├── treesearch/                  # QWBE core, parallel agents, OpenHands coder
    └── utils/

---

Related repositories

• CamylaNet — segmentation framework built on nnU-Net v2, shipping a curated set of CNN / Transformer / state-space backbones. Camyla's baseline stage runs CamylaNet trainers.
• nnPrep — LLM agent that converts arbitrary medical segmentation datasets into the nnU-Net v2 format consumed by CamylaNet (and therefore by Camyla).

---

Limitations & notes

• The baseline stage currently uses CamylaNet trainers. Swapping in another baseline framework requires a corresponding skill under skills/frameworks/.
• OpenHands runs code in your local Python env — make sure you point experiment.openhands.python_path at an env with the needed packages.
• Long runs: a full idea (proposals → 3 stages → paper) typically takes several hours to a day on a single A100-class GPU.

---

Citation

If you find this project useful in academic work, please cite:

@misc{gao2026camyla,
  title         = {Camyla: Scaling Autonomous Research in Medical Image Segmentation},
  author        = {Gao, Yifan and Li, Haoyue and Yuan, Feng and Gao, Xin and Huang, Weiran and Wang, Xiaosong},
  year          = {2026},
  eprint        = {2604.10696},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI}
}

---

Acknowledgements

Camyla builds on ideas and code from two upstream projects:

• AI Scientist (Sakana AI) — pioneered the autonomous research-agent paradigm that inspired Camyla's overall pipeline design.
• nnU-Net — the self-configuring segmentation framework that Camyla's baseline stage (via CamylaNet) is built on.

---

License

Released under the Apache License, Version 2.0 — see LICENSE.