_____ _____ _____ _____ _____ _____ _____ _____ _____     ___ ____  
|   __|   __|   __|   __|     | __  |     |   __| __  |___|_  |    \ 
|__   |   __|  |  |   __|  |  |    -| | | |   __|    -|___|_  |  |  |
|_____|_____|_____|__|  |_____|__|__|_|_|_|_____|__|__|   |___|____/ 

I - Introduction

This repository implements and improves the existing project of Segformer-3D that you can find here : https://github.com/OSUPCVLab/SegFormer3D.

It is related to the research paper on 3D image segmentation cited here.

II - How to setup

Get the GitHub repository by cloning this repo or you can download a folder archive in Kaggle.

At the project root, create a file named .env and define:

TRAIN_DATA_DIR=/path/to/training/data
PROCESSED_DATA_DIR=/path/to/preprocessed/data
OUTPUT_DIR=/path/to/output/models

These paths specify:

• where your raw training data is located
• where preprocessed data will be saved
• where trained models will be stored

Then, you will need to build the docker using docker compose build. This requires that you already have docker installed. The build can last up to 20min if you don't already have the image vllm/vllm-openai:v0.9.1. And the total size of the docker image is around 40GB. When done building, if you try to build again, this time it is only a matter of seconds.

III - How to use

Now, your image is built and you can use it for different purposes. The base command is docker compose run segformer. But as it is, it will do nothing. You have to pass arguments. Here is the full command docker compose run segformer -h -p -t -i <number> -e <number> -w <latest|name> -g -f <fps>.

• -h : Using this option will display how the other options work.
• -p : Using this option will launch the preprocessing of your data.
• -t : Using this option will start the training of a new model on your data.
• -i <number> : Using this option will run a given model on your data. The argument allows to specify which data is evaluated. If you give -1, the model will use all of your data. If you give a positive integer, the model will use the volume associated with this number.
• -e <number> : Using this option will evaluate a model. The argument allows to specify which data is evaluated. If you give -1, the model will use all of your data. If you give a positive integer, the model will use the volume associated with this number.
• -w <latest|name> : [OPTIONAL] Using this option will select an experiment folder as a model loader. If not given, it will select the official Segformer3D model weights. If latest is given, it will select the last experiment folder created automatically. If an experiment name is given, it will use it as model weights.
• -g : [OPTIONAL] Generate GIF animations during inference. By default, animations are skipped for faster processing.
• -f <fps> : [OPTIONAL] Set frames per second for GIF animations (default: 15). Higher values create faster animations. Only used with -g.
  Here are some examples of usage :
• docker compose run segformer -h : Show the help message.
• docker compose run segformer -p -t : Launch the preprocessing and then start training the model.
• docker compose run segformer -t -e -1 : Train the model and the evaluate it on all the data.
• docker compose run segformer -i 1 -w latest : Infer the model saved in the last experiment on the first volume.
• docker compose run segformer -i -1 : Run inference on all validation cases without generating GIF animations (fast, default behavior).
• docker compose run segformer -i -1 -g -f 20 : Run inference on all cases with GIF animations at 20 fps.
• docker compose run segformer -i 42 -g -w latest : Run inference on case 42 using the latest trained model, generating animations.

IV - Research Paper

@inproceedings{perera2024segformer3d,
  title={SegFormer3D: an Efficient Transformer for 3D Medical Image Segmentation},
  author={Perera, Shehan and Navard, Pouyan and Yilmaz, Alper},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={4981--4988},
  year={2024}
}