AnoSiamAE

An anomaly detection model for anterior eye segment leveraging Self-Supervised Siamese Autoencoders. Code soumission for the 12th OMIA workshop on MICCAI 2025.

Installation guidelines

Requirements

This repo was build upon python 3.9 and all dependencies are given in the requirements.txt file.

  pip install -r requirements.txt

Datasets

We used the RFMiD dataset, a public collection of 3,200 retinal fundus images annotated with 46 pathologies. It includes 1,920 training images (of which only the 400 normal images are used for training), and 640 images each for validation and testing.

After installing the RFMiD dataset, make sure it follows the adequate folder structure:

RFMiD  
|-> Training
|   |-> XXX.png
|   |-> ...
|-> Validation
|    |-> XXX.png
|   |-> ...
|-> Test
|   |-> XXX.png
|   |-> ...
|-> RFMiD_Testing_Labels.csv
|-> RFMiD_Training_Labels.csv
|-> RFMiD_Validation_Labels.csv

By default, the path to the RFMiD dataset is set to

RFMiD = "/store/data/RFMiD"

in the src/utils/constant.py file. Make sure to use the right path according to your dataset installation location. Custom dataset can be used by adding a new dataloaders (src/data/dataloader.py) and datasets (src/data/datasets.py) in their corresponding file. Make sure to follow the same structure as the examples given for the proposed RFMiD dataloaders, and by adding them in the DATALOADER_REGISTRY. Path to the new dataset can be added in the src/data/access_data.py file.

Preprocessing

Data are preprocessed to remove the background and only keep the ROI using domain-specific hyperparameters. Make sure these parameters work with your input data. (src/data/transform/preprocess.py)

Run training

Configuration files

Specific configuration of all baselines models are given in the configuration folder:

config
|-> ae_mae.yaml
|-> ae_ssim.yaml
|-> anovaegan.yaml
|-> beta_vae.yaml
|-> dae.yaml
|-> dpae.yaml
|-> draem.yaml
|-> ganomaly.yaml
|-> siamae.yaml

Global training and evaluation configuration parameters, including data augmentations and system parameters, are set in the src/config/config.py file.

Training

Models can be trained using the main.py script following the given prompt:

python main.py --cfg_path /path/to/cfg.yaml 

Afterwards, training will start and a folder in output/default/model_name will be created, where logs, visulization steps and model weights will be saved.

Several flag options are made available for training:

--cfg_path (str): Path to model configuration file
--ckpt_path (str): Path to model checkpoint path (continue training from saved checkpoint)
--folder (str): Path to model output folder (continue training from saved output model)
--output (str): Override configuration output path to the given path
--data (str): Data name (default:rfmid)
--device (str): cuda/cpu (default:cuda)
--batch_size (int):Number of samples per batch

Evaluation

Models can be tested using one of the following prompts:

python main.py --cfg_path /path/to/cfg.yaml --ckpt /path/to/checkpoint.pth --test_only

python main.py --folder /path/to/model/output/folder --test_only

Cite

@InProceedings{10.1007/978-3-032-10351-2_5,
  author="Baumstimler, Philippe
  and Gagn{\'e}, S{\'e}bastien
  and Champagne, Mathieu
  and Gagnon, Jean-Mathieu
  and Playout, Cl{\'e}ment
  and S{\'e}oud, Lama",
  editor="Fang, Huihui
  and Wang, Meng
  and Li, Heng
  and Chen, Hao
  and Bogunovi{\'{c}}, Hrvoje
  and Lee, Cecilia S.",
  title="Anomaly Detection in Anterior Eye Segment Using Self-supervised Siamese Autoencoders",
  booktitle="Ophthalmic Medical Image Analysis",
  year="2026",
  publisher="Springer Nature Switzerland",
  address="Cham",
  pages="43--53",
}