Facebook's DETR object detection

Automatic localization and classification of thoracic abnormalities from chest radiographs for the VinBigData Chest X-ray Abnormalities Detection Kaggle competition using Facebook DETR's detection.

Input Datasets:

A) Image dataset: (test and train folder)
18000 images (train set: 15000 images, test set: 3000 images)

The image dataset includes 14 types of thoracic abnormalities from chest radiographs as well as images with no abnormality detected (no finding). The different labels are as follows:
0 - Aortic enlargement
1 - Atelectasis
2 - Calcification
3 - Cardiomegaly
4 - Consolidation
5 - ILD
6 - Infiltration
7 - Lung Opacity
8 - Nodule/Mass
9 - Other lesion
10 - Pleural effusion
11 - Pleural thickening
12 - Pneumothorax
13 - Pulmonary fibrosis
14 - No finding

B) Image metadata (train.csv)
the train set metadata, each row represents an abnormality of one image, its class and bounding box. Some images can contain multiple abnormalities.

Prerequisites:

Running the DETR model requires the use of a GPU and Kaggle's notebook environment can provide that.
The code of this repository can be also found in this Kaggle's public notebook.

For the training:

1. Select the mode = 'train'
   
2. Add data --> Competitions Data --> Search for "VinBigData Chest X-ray Abnormalities Detection"
   
3. Add data --> Datasets --> Search for "vinbigdata-chest-xray-original-png"
   
4. Enable the GPU in the Settings --> Accelarator --> GPU
   The output of the mode is "detr_model.pth"
   

For the predictions:

1. Select the mode = 'predict'
   
2. Go to the outputs of the previous version (train mode "detr_model.pth"), select "New dataset" and keep the created URL
   
3. Go back to Kaggle's notebook --> Add data --> Datasets --> Search by URL with the saved URL from the last step
   
4. Εnable the CPU in the Settings --> Accelarator --> CPU
   

Results:

Learning curve

Mean average precision