Skip to content

Latest commit

 

History

History
61 lines (51 loc) · 2.91 KB

README.md

File metadata and controls

61 lines (51 loc) · 2.91 KB
Raw

Pneumothorax-Detection

Repository for code tested to build models for the Kaggle SIIM-ACR Pneumothorax Segmentation contest.

Managing DICOM images: Tips and tricks for the radiologist https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3354356/

CheXNeXt PLOS article https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1002686

CheXNet: Radiologist-Level Pneumonia Detection | Kaggle https://www.kaggle.com/ashishpatel26/chexnet-radiologist-level-pneumonia-detection

Original CheXNet arxiv article https://arxiv.org/abs/1711.05225

Chester: A Web Delivered Locally Computed Chest X-Ray Disease Prediction System https://arxiv.org/abs/1901.11210

Google AI algorithm may improve chest x-ray interpretation, radiologist efficiency https://www.healthimaging.com/topics/artificial-intelligence/google-ai-algorithm-may-improve-chest-x-ray-interpretation

Thoracic Disease Identification and Localization with Limited Supervision https://arxiv.org/abs/1711.06373

Pneumothorax detection with CheXNet:

  • A binary classification problem.
  • CheXNet is a 121-layer Dense Convolutional Network (DenseNet) trained on the ChestX-ray 14 dataset.
  • DenseNets improve flow of information and gradients through the network, making the optimization of very deep networks tractable.
  • The final fully connected layer is replaced with one that has a single output, after which we apply a sigmoid nonlinearity. (we can consider using a final SVM layer?)
  • The weights of the network are initialized with weights from a model pre-trained on ImageNet (Deng et al., 2009).
  • The network is trained end-to-end using Adam with standard parameters (β1 = 0.9 and β2 = 0.999) (Kingma & Ba, 2014).
  • The model was trained using mini batches of size 16.
  • An initial learning rate of 0.001 that is decayed by a factor of 10 each time the validation loss plateaus after an epoch.
  • The model with the lowest validation loss was picked.
  • Can add to our detection mechanism Sobel edge detection or operator(s): https://www.sciencedirect.com/topics/engineering/sobel-operator
  • Assume no patient overlap between the training and testing datasets.
  • Pathologies covered:
    1. Atelectasis
    2. Cardiomegaly
    3. Effusion
    4. Infiltration
    5. Mass
    6. Nodule
    7. Pneumonia
    8. Pneumothorax
    9. Consolidation
    10. Edema
    11. Emphysema
    12. Fibrosis
    13. Pleural Thickening
    14. Hernia
  • CheXNet detection accuracy = 0.8887 (2017)
  • Possible use case for NASNet. Currently inserted Keras' default NASNET into the model. Have a look at the architectures mentioned here: https://towardsdatascience.com/illustrated-efficient-neural-architecture-search-5f7387f9fb6#df67
  • Will look into Reinforcement Learning for an effective implementation of ENAS.

UNet Tutorials