DETR LUNA

Fork of Facebook's DETR Object Detection with Transformers model using the LUNA16 lung cancer dataset.

Description

This fork uses 2D CT scan images extracted from the LUNA16 lung cancer dataset. The original LUNA16 dataset contains 3D CT scan objects. For this work, these objects were converted to 2D 512x512 images. Some data-augmentation techniques were used, giving us around 4,000 images with one of more lung nodules.

Configuring the environment

1. Download the images and annotations from this link and extract them to the dataset/ directory, in the following structure:

   luna_images_test/
   luna_images_train/
   luna_images_seg_test/
   luna_images_seg_train/
   test.json
   test_seg.json
   train.json
   train_seg.json

2. Download the COCO images pretrained weights to detr/weights/.

3. Create a conda environment using python 3.7 and install the requirements:

   conda create -n DETR python=3.7
   conda activate DETR
   pip install -r detr/requirements.txt

Training the model

Run the main.py code from within the detr directory:

python main.py --device cuda --dataset_file luna --data_path ../dataset/ --output_dir output_2/ --resume weights/detr-r50-dc5-f0fb7ef5.pth --batch_size 8 --lr_drop 1000 --num_queries 10 --epochs 600

We start the training using the pre-trained weights given in the repo. After training our model, we can continue further epochs refering to our own checkpoint file as --resume output/checkpoint.pth.