Retinograd ai

Gradeability model for FAF images.

This repository contains a Python script for performing image classification using a pre-trained Inception ResNet v2 model specifically designed to assess the gradeability of Fundus Autofluorescence (FAF) images. The script reads image paths from a CSV file, processes each image, and outputs the gradeability predictions to a new CSV file. This tool is intended to assist researchers and clinicians in evaluating the quality and diagnostic utility of FAF images.

Note

If you use Retinograd in your work, please cite us as follows:

Retinograd-AI: An Open-source Automated Gradeability Assessment for Retinal Scans for Inherited Eye Dystrophies
URL: https://doi.org/10.1101/2024.08.07.24311607
_{^{William Woof*, Gunjan Naik*, Saoud Al-Khuzaei*, Thales Antonio Cabral De Guimaraes, Malena Daich Varela, Sagnik Sen, Pallavi Bagga, Ismail Moghul, Michel Michaelides, Konstantinos Balaskas, Nikolas Pontikos}}

Setup

1. Clone the repository:

   git clone https://github.com/Eye2Gene/retinograd-ai.git
   cd retinograd-ai

2. Install the required packages

   pip install -r requirements.txt

3. Download model weights

   Go to the Releases page of this repository. Under the latest release, find the FAF_inception_resenet_2Class_classification.pth file listed under "Assets" and download it to the weights directory in your project.

   wget -O ./weights/model.pth https://github.com/Eye2Gene/retinograd-ai/releases/download/V0.0.1/FAF_inception_resenet_2Class_classification.pth

4. Train or Finetune model:

   For training of model:

   If you want to train the model from scratch for your custom data, please run below command.

    python  train.py --mode train --train_csv <csv_path_of_train> --val_csv <csv_path_of_val> --image_path_csv_column <image_path_column> --label_csv_column <label_column> --model_path <path_to_save_model_weights> --num_epochs <number of epochs> --batch_size <number_of_images_per_batch> --learning_rate <learning_rate>

   For Finetuning of model:

   If you want to Finetune the model using our model weights for your custom data, please run below command. For this, the model weights shoule be downloaded first.

    python  train.py --mode finetune --train_csv <csv_path_of_train> --val_csv <csv_path_of_val> --image_path_csv_column <image_path_column> --label_csv_column <label_column> --model_path <path_to_save_model_weights> --num_epochs <number of epochs> --batch_size <number_of_images_per_batch> --learning_rate <learning_rate>v
   

   The train and val csv should have one column with input image path and second column for labels. The labels columns should have data type 'int' and have values 0 and 1 only. The class '0' is for Ungradable cases and class '1' is for Gradable cases.

5. Running Inference

   python classify_images.py --csv_path <input_csv> --csv_column <image_path_column> --model_path <model_weights> --output_csv_path <output_csv>
   
   # For example:
   python classify_images.py --csv_path data/images.csv --csv_column image_path --model_path weights/model.pth --output_csv_path data/predictions.csv

Train/Finetune/Inference Arguments

• --csv_path (str): Path to the input CSV file containing image paths.
• --csv_column (str): Column name in the CSV file that contains the image paths. Default is image_path.
• --model_path (str): Path to the model weights file. Model weight file has been added in releases.
• --output_csv_path (str): Path to save the output CSV file with predictions. Default is predictions.csv.

The input CSV requires at least one column containing all the paths to the input images. The name of this column should be passed using --csv_column <col_name>.

Inference Outputs

Retinograd-ai currently predicted two labels for each image:

1. Class 0: Ungradable
2. Class 1: Gradable

License

This project is licensed under the MIT License. See the LICENSE file for details.