glaucoma-rdcnn

Replication of Region-based Deep CNN (R-DCNN) for joint optic disc and optic cup segmentation, from Li, Xiang, Zhang et al., "Joint optic disk and cup segmentation for glaucoma screening using a region-based deep learning network", Eye 37, 1080–1087 (2023) — DOI 10.1038/s41433-022-02055-w.

R-DCNN reformulates joint optic-disc/optic-cup (OD/OC) segmentation as object detection: a Faster-R-CNN-style network predicts an OD bounding box, a second head conditioned on the OD region predicts the OC bounding box, and an inscribed ellipse is fit inside each box to produce the final segmentation. Cup-to-disc ratio (CDR) follows directly from the two ellipses; CDR > 0.5 is the standard glaucoma-suspect threshold.

Status. Scaffold + reference implementation. Numbers in the "this repo" column below are intentionally blank until full training is run on Spartan — see the open issues for the reproduction tracker.

Architecture

RGB fundus  ──► CLAHE ──► OD localizer ──► 800x800 ROI crop
                                              │
                                              ▼
                       ┌────────────────────────────────────┐
                       │  ResNet-34 (ImageNet) + DAC block  │ ← shared backbone
                       └────────────────────────────────────┘
                                  │
                ┌─────────────────┴─────────────────┐
                ▼                                   ▼
       ┌─────────────────┐    Disc Attention    ┌─────────────────┐
       │  DPN (RPN+ROI)  │ ─────────────────►  │  CPN (RPN+ROI)  │
       │  → OD bbox      │                     │  → OC bbox      │
       └─────────────────┘                     └─────────────────┘
                │                                   │
                └─────────────► inscribed ellipse fitting ◄────┘
                                          │
                                          ▼
                                 OD mask, OC mask, CDR

Components:

• Backbone: ResNet-34 (torchvision, ImageNet pretrained) with a Dense Atrous Convolution (DAC) block at the deepest stage to widen receptive field without losing resolution. See src/glaucoma_rdcnn/models/dac.py.
• Disc Proposal Network (DPN): anchor-based RPN + ROI head producing the OD bounding box. Built on torchvision.models.detection.rpn to avoid reimplementing the proposal machinery.
• Cup Proposal Network (CPN): same structure as DPN, but receives features that have been re-attended through the predicted OD region (Disc Attention Module).
• Geometry: the largest ellipse inscribed in each predicted axis-aligned box. CDR = vertical-cup-diameter / vertical-disc-diameter.

Reproduction targets

Numbers from the paper (Table 1 + Fig 3) — populate the "this repo" columns after a real training run:

Dataset	OD DC (paper)	OD DC (ours)	OD JC (paper)	OD JC (ours)	OC DC (paper)	OC DC (ours)	OC JC (paper)	OC JC (ours)	Glaucoma AUC (paper)	AUC (ours)
DRISHTI-GS	97.23%		94.17%		94.56%		89.92%		0.968
RIM-ONE v3	96.89%		91.32%		88.94%		78.21%		0.941

Install

git clone https://github.com/basim-azam/glaucoma-rdcnn.git
cd glaucoma-rdcnn
python -m venv .venv && source .venv/bin/activate    # or use conda
pip install --upgrade pip
# CPU dev: torch from CPU index
pip install torch==2.5.1 torchvision==0.20.1 --index-url https://download.pytorch.org/whl/cpu
pip install -e ".[dev]"
pre-commit install
pytest -q

For GPU on Spartan, see docs/spartan_guide.md. The conda env is built by slurm/001_setup_env.slurm at /data/gpfs/projects/punim2920/glaucoma-rdcnn/conda_envs/glaucoma.

Datasets

Two public datasets are auto-downloaded; a third (in-house) is documented as a pluggable schema.

Dataset	Images	OC mask	Notes
DRISHTI-GS	101	yes	https://cvit.iiit.ac.in/projects/mip/drishti-gs/mip-dataset2/Home.php
RIM-ONE v3	159	yes	Public mirror referenced in paper
Custom (in-house)	—	yes	Plug your data in via configs/data/custom.yaml — see docs/dataset_schema.md

bash scripts/download_data.sh         # ~250 MB, idempotent, sha256-verified
python scripts/preprocess.py --config-name=data/drishti_gs

Quickstart

# Smoke training (one batch, CPU)
python scripts/train.py --config-name=experiment/repro_drishti trainer.fast_dev_run=true

# Real training on a single GPU
python scripts/train.py --config-name=experiment/repro_drishti training=single_gpu

# Evaluate
python scripts/evaluate.py --config-name=experiment/repro_drishti checkpoint=outputs/best.ckpt

# Inference on a single fundus image
python scripts/infer.py --image=path/to/fundus.jpg --checkpoint=outputs/best.ckpt

Spartan HPC

This repo is wired for the University of Melbourne Spartan cluster, project allocation punim2920. The slurm/ directory has scripts for env setup, data download, smoke test, single/multi-GPU training, and evaluation. See docs/spartan_guide.md.

ssh spartan
cd /data/gpfs/projects/punim2920
git clone https://github.com/basim-azam/glaucoma-rdcnn.git
cd glaucoma-rdcnn
sbatch slurm/001_setup_env.slurm    # one-time
sbatch slurm/003_smoke_test.slurm   # gpu-a100-short, ~30 min
sbatch slurm/004_train_a100_1gpu.slurm

Citation

If you use this code, please cite both the original paper and this repository:

@article{li2023joint,
  title   = {Joint optic disk and cup segmentation for glaucoma screening using a region-based deep learning network},
  author  = {Li, Feng and Xiang, Wei and Zhang, Lin and Pan, Wenjuan and Zhang, Xuemin and Jiang, Mingshuai and Zou, Haidong},
  journal = {Eye},
  volume  = {37},
  pages   = {1080--1087},
  year    = {2023},
  doi     = {10.1038/s41433-022-02055-w}
}

@software{azam2026rdcnn,
  author = {Basim Azam},
  title  = {glaucoma-rdcnn: Replication of R-DCNN for joint OD/OC segmentation},
  year   = {2026},
  url    = {https://github.com/basim-azam/glaucoma-rdcnn}
}

License

MIT — see LICENSE.