KoSAIM2022-chest-x-ray-classification

[KoSAIM 2022 Summer School] Hands-On-CNN: Chest X-ray Classification

Colab Link

Slides

Google Drive Dataset

objective

Given a chest X-ray image, you need to predict positive diagnosis (pneumonia) labels. Construct the binary image classifier for prediction, and you could consider any kind of image encoder.

dataset

Ped-Pneumonia
https://www.kaggle.com/datasets/paultimothymooney/chest-xray-pneumonia
http://www.cell.com/cell/fulltext/S0092-8674(18)30154-5

The normal chest X-ray (left panel) depicts clear lungs without any areas of abnormal opacification in the image. Bacterial pneumonia (middle) typically exhibits a focal lobar consolidation, in this case in the right upper lobe (white arrows), whereas viral pneumonia (right) manifests with a more diffuse ‘‘interstitial’’ pattern in both lungs.

images & labels

Feel free to pre-process the images as you see fit (the minimum resolution is 224x224).

things to consider

• Image resolution
• Pixel value normalization
• Class imbalance

data split

modeling & evaluation

• Use whatever model you prefer.
  • Recommended: any CNN-based model (e.g., VGGNet, ResNet)
• Use both AUROC and AUPRC as the evaluation metric.