top of page
Covid-19 Prediction

Building a Deep Learning Model Uses CT Images for Covid-19 Diagnosis

  • Lung CT images which are taken from Tongji Hospital, Wuhan, China for January 2020, and April 2020 are analyzed

  • The whole image dataset consists of 349 CT images from 296 patients diagnosed with Covid-19 and 397 CT images from non-covid patients

  • The images are resized to train the model faster as 100x100 and converted into torch tensors

  • The whole dataset is splitted into train, validation, and test datasets

Ekran Alıntısı4.PNG
Ekran Alıntısı.PNG
Adsız.png

  • A convolutional neural network with 3 convolutional and 3 pooling layers, and a final fully connected layer is used

  • Batch normalization after convolutional layers are applied to speed up training

  • ReLU activation function is used after hidden layers to eliminate vanishing gradient problem

  • Softmax activation function in final layer to turn the vector into probabilities that sum up to 1

  • Cross entropy loss function is chosen since it is wiser to use it in classification problems

  • Adam and SGD optimizers are used

Neural Style Art Transfer

  • Neural Style algorithm developed by Leon A. Gatsy, Alexander S.Ecker and Matthias Bethge is applied

  • Neural style is creating an image which has the content of an image and a style of another image

  • The key technique used in neural style transfer is convolutional neural network

  • Two different pictures are taken from the internet and required preprocessing steps are taken
    • They are resized to same shape as 512x512
    • Converted into tensors

  • The loss functions for both content and style are defined

  • 19-layer VGG network is used
  • Adam and SGD optimizers are both applied, and it is observed that results by Adam optimizer are better
  • After performing neural transfer with different learning rates, number of epochs, and weights for content and style, final output which is the combination of 2 input images is obtained
  • Since the quality of an image is subjective, it is hard to evaluate the performance of the neural style transfer. Thus, the results are not interpreted in a quantitative manner
Ekran Alıntısı.PNG
dancing.jpg
Ekran Alıntısı.PNG

Chest X-Ray Pneumonia Classification

  • 5856 chest X-ray images are used

  • Three research questions which are tried to be answered:
    • Do different networks result similar/same performance?

    • Does applying transfer learning influence performance?

    • Does using deep networks without pretraining on ImageNet have similar performance with the pretrained version?

  • Data preprocessing methods are conducted to have better results in further modeling studies.

  • Different models are run and finally, the models are evaluated with appropriate metrics and the results are interpreted accordingly

  • Future works:
    • To use a source dataset which is in the similar domain with the target dataset during transfer learning
    • To increase the number of images is data augmentation
    • To train the deep networks partially
deepnetworkds.JPG
xray.JPG
heatmap.JPG
Neural Style Art Transfe
Pneumonia Classification
bottom of page