This repository provides training and inferencing modules for Pneumonia Detection. This aim of this project is to provide a cookie cutter approach to generate bounding box predictions for Chest X Rays.
There are many object detection techniques are available today for generation of bounding boxes like YoloV5, R-CNN, Faster R-CNN etc. Segmentation is primarily a technique which marks. Here I am using segmentation to generate bounding boxes.
This project provides two functions:
- Inferencing on Chest XRay Images to find Bounding box and associated labels. (A pretrained model is provided, but users are encouraged to train their own models based on the training script provided as below.
- Training of model : The data setup is explained and code is made available in the Dataset up section below. The data for traning is required in two folders names images and masks under a parent folder, say train. At time of training, the location of this folder (train) needs to be passed to the training script (train.py).
BIMCV COVID-19+ (https://www.kaggle.com/c/siim-covid19-detection/data).
The bounding box and label information is provided in csv files. For segmentation, masked images are generated from the given information which is used for model generation.
Following notebooks contain the code for data set up.
- zip-to-dcm.ipynb - Extracts dicom files from zip files.
- dcm-to-png.ipynb - Extracts png files from dicom files. Code for cleaning the corrupted images is also available.
- png-to-mask.ipynb - Creates mask images for corresponding X Rays using information provided in excel files.
Get the repository: git clone https://github.com/sinharitesh/bounding-box-using-segmentation.git
Change the directory: cd bounding-box-by-segmentation
Create a directory "models" in current folder: mkdir models
pip install -r requirements.txt
Change directory to models: cd models
Download a pre trained model provided for this purpose using following in the models directory :
pip install gdown
gdown https://drive.google.com/uc?id=1k6jTUj5p7WzqcE8B11ha6rsRiR3j8H-A
Above will download a model named "siim-seg-011-resnet34-colab.pth".
Comeback to root directory (one directory up)
Once you are in the directory containing the script train.py, run the following command:
python train.py --train_dir ../data-siim/train/ --use_model siim-seg-011-resnet34-colab.pth --epochs 1 --learning_rate .01
Training will start, after the training is finished, a model named siim-segmentation-resnet34-.pth will be available in models directory.
Note: For next training session, above saved model can be passed and the training will be resumed on latest saved model.
Put the images which need to be inferenced in the folder "test/images".
Some images are made available as samples. There is also a folder called labels under the folder test which contains ground truth information for images. This is from the perspective to visually examine the prediction done on the images. This is not a requirement for inferencing.
Run the following command:
python inference.py --test_dir ./test/
Note that you can pass any directory which has got images subfolder containing the images to be examined.
After inferencing is completed, the output as images with bounding box drawn over them is saved in the folder called outputs.
The name of the output file will be sample001_pred.png where original file name was sample001.png. In case of ground truth information is also available, there will be an additional file named as sample001_orig.png
The architecture used here is unet, with backbone resnet34. Fast.ai library is used for building models which utilizes Pytorch.
Fast.ai has inbuilt facility to find the optimal learning rate, one can use the following notebook to determine learning rate.
source file: learning-rate-finder.ipynb
The BIMCV-COVID19 Data used by this challenge were originally published by the Medical Imaging Databank of the Valencia Region (BIMCV) in cooperation with The Foundation for the Promotion of Health and Biomedical Research of Valencia Region (FISABIO), and the Regional Ministry of Innovation, Universities, Science and Digital Society (Generalitat Valenciana), however the images were completely re-annotated using different annotation types. Users of this data must abide by the BIMCV-COVID19 Dataset research Use Agreement. Paper Reference: BIMCV COVID-19+: a large annotated dataset of RX and CT images from COVID-19 patients
The MIDRC-RICORD Data used by this challenge were originally published by The Cancer Imaging Archive. The images were re-annotated for this challenge using a different annotation schema. Users of this data must abide by the TCIA Data Usage Policy and the Creative Commons Attribution-NonCommercial 4.0 International License under which it has been published. Attribution should include references to citations listed on the TCIA citation information page (page bottom). Paper Reference: The RSNA International COVID-19 Open Radiology Database (RICORD)