Xvision

Chest Xray image analysis using Deep Learning and exploiting Deep Transfer Learning technique for it with Tensorflow.

The maxpool-5 layer of a pretrained VGGNet-16(Deep Convolutional Neural Network) model has been used as the feature extractor here and then further trained on a 2-layer Deep neural network with SGD optimizer and Batch Normalization for classification of Normal vs Nodular Chest Xray Images.

Nodular vs Normal Chest Xray

<img src="https://github.com/ayush1997/Xvision/blob/master/image/node.jpg" width="300" height="300" /> <img src="https://github.com/ayush1997/Xvision/blob/master/image/normal.jpg" width="300" height="300" />

Some specifications

| Property |Values | | ------------- | ------------- | | Pretrained Model | VggNet-16 | | Optimizer used | stochastic gradient descent(SGD) | | Learning rate | 0.01|
|Mini Batch Size| 20 | | Epochs | 20 | |2 Layers| 512x512 | |GPU trained on| Nvidia GEFORCE 920M|

Evaluation

Confusion Matrix and Training Error Graph

<img src="https://github.com/ayush1997/Xvision/blob/master/image/cfm.jpg" width="450" height="400" /> <img src="https://github.com/ayush1997/Xvision/blob/master/image/nodule.jpg" width="400" height="400" />

| | Normal | Nodule | |------|---------|---------| | Precision| 0.7755102| 0.55555556 | |Recall| 0.76 | 0.57692308 |

Accuracy : 69.3333 %

DataSet

openi.nlm.nih.gov has a large base of Xray,MRI, CT scan images publically available.Specifically Chest Xray Images have been scraped, Normal and Nodule labbeled images are futher extrated for this task.

How to use ?

The above code can be used for Deep Transfer Learning on any Image dataset to train using VggNet as the PreTrained network.

Steps to follow

1. Download Data- the script download images and saves corresponding disease label in json format.

python scraper.py <path/to/folder/to/save/images>

2. Follow the scraper/process.ipynb notebook for Data processing and generate

• Training images folder - All images for training
• Testing images Folder - All images for testing
• Training image labels file - Pickled file with training labels
• Testing image labels file - Pickled file with testing labels

3. Extract features(CNN Codes) from the maxpool:5 layer of PreTrained CovNet(VggNet) and save them beforehand for faster training of Neural network.

   python train.py <Training images folder> <Testing image folder> <Train images codes folder > <Test images codes folder>

   • Train images codes folder - Path where training images codes will be stored
   • Test images codes folder - Path where testing images codes will be stored

4. The extracted features are now used for training our 2-Layer Neural Network from scratch.The computed models are saved as tensorflow checkpoint after every Epoch.

   python train_model.py <Training images folder> <Train images codes folder> <Training image labels file> <Folder to save models>

5. Finally the saved models are used for making predictions.Confusion Matrix is used as the Performance Metrics for this classifcation task.

   python test_model.py <Testing images folder> <Test images codes folder> <Testing image labels file> <Folder with saved models>

Some Predictions

References

1. Learning to Read Chest X-Rays: Recurrent Neural Cascade Model for Automated Image Annotation

2. Deep Convolutional Neural Networks for Computer-Aided Detection: CNN Architectures, Dataset Characteristics and Transfer Learning

Contribute

If you want to contribute and add new feature feel free to send Pull request here :D

To report any bugs or request new features, head over to the Issues page

To-do

• [ ] Implement saliency map or use Deconv for better visualizations.