SIIM-FISABIO-RSNA-COVID-19-Detection

Details of the Competition

• Objective : Identify and localize COVID-19 abnormalities on chest radiographs.
• Host : Society for Imaging Informatics in Medicine (SIIM)
• Partners : HP & Intel
• Website : Kaggle
• Timeline : May 18,2021 -> August 10,2021
• Evaluation Criteria : mean Average Precision(mAP) at IoU > 0.5

Dataset

Data Structure

• train folder (contains 6300 chest scans in DICOM format)
  • study
    • series
      • image
        • .dicom files
• test data (contains 1200 chest scans in DICOM format)
  • study
    • series
      • image
        • .dicom files
• sample_submission.csv
• train_image_level.csv
  • id - unique image identifier
  • boxes - bounding boxes in easily-readable dictionary format
  • label - the correct prediction label for the provided bounding boxes
• train_study_level.csv
  • id - unique study identifier
  • Negative for Pneumonia - 1 if the study is negative for pneumonia, 0 otherwise
  • Typical Appearance - 1 if the study has this appearance, 0 otherwise
  • Indeterminate Appearance - 1 if the study has this appearance, 0 otherwise
  • Atypical Appearance - 1 if the study has this appearance, 0 otherwise

Data Preprocessing

• DICOM images are converted to JPG images with various sizes (224,256,512,1024) using the following Data prepartion notebook.
• Some of the metadata of train and test data is also stored in csv files.
• Dataset is uploaded in Kaggle.

Data Visualization

• To understand the data more clearly, data visualization is made using seaborn, matplotlib, Pandas Profiling wordcloud.
• Some of the Plots :

Training

Study Level

• Study Level prediction is a multi class classification task i.e., we have to predict whether given chest x-ray belongs to one of the Negative for Pneumonia, Typical Appearance, Indeterminate Appearance, Atypical Appearance categories.
• TensorFlow pretrained models, ChexNet model, various image sizes are used for experimentation.
• More details are in Study Level Prediction notebook.
• Best models are saved in .h5 format.

Image Level

• Image Level prediction is a object detection task where we have to localize the abnormality in chest x-rays.
• I used YOLOv5 to train the model using various image sizes and cross validation techniques.
• More details of training are in Image Level Prediction notebook.
• Best models are saved in .pt format.
• Training results are examined using Wieghts & Biases dashboard.
• To know more about Object detection and it's evolution over years, read research papers present in papers folder.

Inference

• Finally the saved models from Study Level and Image Level are loaded and used them predict result on unseen test data which contains over 1200 chest x-rays.
• More details of Inference are in Inference notebook.