FYP Plan (Trying To Break Things)````

Methodology

Currently, stage 1 StyleGAN is trained using the CheXpert dataset. This dataset consists of around 200k chest X-rays consisting of 14 different labels: chexpert_labels.png

To verify that stage 2 works, we have to ensure that:

  • Stage 1 StyleGAN is capable of generating CXR images with the intended pathology
  • Find CT scans with the desired pathology and check that generated CXR actually contains the pathology (or even more details)

More details on each pathology:

  • Enlarged Cardiom: not sure, to find out
  • Cardiomegaly: Enlarged heart. Can diagnose in CXR
  • Lung Lesion, Lung Opactiy, Edema, Consolidation: generic observations, not very valuable, need CT scan to diagnose more specific conditions
  • Pneumonia: Pneumonia, not very certain in CXR
  • Atelectasis: Lung collapse, not too sure in CXR
  • Pneumothorax: Leaky lung, quite certain in CXR

So the plan is: Stage 1 StyleGAN Training

  • Select subset of CheXpert

Ideas:

  • StyleGAN increase higher resolution channels, reduce lower resolution channel

ESP32-C3 > Documents https://github.com/rvignav/CT2Xray