OCTA-SVR (Synthetic Vasculature Reasoning)

Public release of the core code used in the Synthetic Vasculature Reasoning (SVR) paper: pathology-aware OCTA synthesis, automatic reasoning text generation, and minimal LLaMA-Factory configs for training/inference.

• Pathology-aware synthesis and VLM pair creation are under pathology/ (compatible with OCTA-autosegmentation).
• LLaMA-Factory assets are under llama_factory/ (upstream repo: https://github.com/hiyouga/LLaMA-Factory).
• OCTA-100K-SVR dataset: https://zenodo.org/records/17706653

Quickstart

git clone https://github.com/d0ng231/OCTA-SVR
cd OCTA-SVR
python3 -m venv .venv  # Python >= 3.10 recommended
source .venv/bin/activate
pip install -r requirements.txt

Synthetic pathology generation

1) Standard pathology-aware dataset (dropout + MA + NV)

To generate a pathology-aware OCTA-style dataset with capillary dropout, microaneurysms, neovascular tufts and mild tortuosity, use the provided config:

# Graphs + vessel maps + overlays (no GAN, fast)
python pathology/generate_vlm_dataset.py \
  --config_file pathology/generation_vlm_dropout_ma.yml \
  --stage graphs

# Full pipeline with GAN + image–text pairs
python pathology/generate_vlm_dataset.py \
  --config_file pathology/generation_vlm_dropout_ma.yml \
  --stage all \
  --use_gan \
  --gan_config /path/to/OCTA-autosegmentation/docker/trained_models/GAN/config.yml \
  --gan_epoch 150 \
  --gan_device cuda

Outputs (under data/vlm_dataset_dropout_ma/ by default):

• vessel_graphs/**/: CSV graphs, art_ven_img_gray.png, pathology_overlay_white.png, pathology_vis.png, dropout_ma.json, pathology.yml.
• images/M_*.png: panned/cropped vessel maps with pathology overlays (optionally GAN-based OCTA if --use_gan is enabled).
• metadata.jsonl, pairs.jsonl: reasoning-ready VLM data (ShareGPT-style JSONL).

The YAML pathology/generation_vlm_dropout_ma.yml is a simplified port of configs/generation_vlm_dropout_ma.yml from OCTA-autosegmentation, keeping only the knobs needed for SVR.

2) Low-level vessel graph + vessel map demo (no pathology)

python pathology/generate_synthetic_octa_images.py \
  --num_samples 100 \
  --out_dir data/vlm_dataset_demo \
  --workers -1

This creates per-sample graph folders under data/vlm_dataset_demo/vessel_graphs/ that contain CSV graphs and grayscale vessel maps (art_ven_img_gray.png).

To additionally render realistic OCTA images with the GAN from OCTA-autosegmentation, append:

  --use_gan \
  --gan_config /path/to/gan/config.yml \
  --gan_epoch 150

GAN configs/weights (and the test.py runner) live in the upstream OCTA-autosegmentation repo; to use GAN rendering, run this script inside that repo or copy it into that workspace. In most SVR use cases you do not need to call this script directly—prefer the config-driven generate_vlm_dataset.py pipeline above.

3) Generate VLM text pairs

If you already have graphs/images under a dataset root, you can regenerate only the VLM text pairs (without re-running growth or GAN) via:

python pathology/generate_vlm_dataset.py \
  --config_file pathology/generation_vlm_dropout_ma.yml \
  --out_dir data/vlm_dataset_dropout_ma \
  --stage pairs

This reads existing vessel_graphs/ and images/ under data/vlm_dataset_dropout_ma/ and writes/updates metadata.jsonl and pairs.jsonl (ShareGPT-style conversations) in that folder.

Key outputs for any dataset root: images/*.png, metadata.jsonl, and pairs.jsonl. Pathology ranges and profiles are best adjusted via the YAML configs in pathology/ rather than editing Python.

Training with LLaMA-Factory

Place OCTA-100K-SVR under data/OCTA-100K-SVR/ (so pairs_diversified.jsonl and images/ live there), then run:

bash llama_factory/scripts/run_svr_training.sh

The script calls llamafactory-cli with llama_factory/configs/svr_qwen3vl_full.yaml and llama_factory/data/dataset_info.json. Edit the config to swap the base model or adjust batch size/epochs. Refer to the upstream LLaMA-Factory docs for environment/setup details.

Inference

# Single image
bash llama_factory/scripts/run_inference.sh \
  MODEL_PATH=/path/to/qwen3vl_svr_checkpoint \
  IMAGE_PATH=/path/to/image.png \
  OUT_PATH=outputs/predictions.jsonl

The underlying helper llama_factory/inference_octa_CoT.py supports both single-image and folder inference, optional LoRA adapters, and template selection (qwen3_vl, qwen2_vl, etc.).

Repo layout

• pathology/: Synthetic vessel/pathology simulation, CoT pairing, cleanup.
• pathology/vessel_graph_generation/: Graph-growth simulator and rasterization.
• llama_factory/: Training config, dataset descriptor, inference helpers, scripts.
• requirements.txt: Minimal deps (GAN weights/models not included).