(port and enhancement of tensorflow repo found at ayush1997/Xvision using Resnet-50)
Chest X-Ray image analysis using Deep Transfer Learning technique. Written in python for MxNet deeplearning framework.
Summary: The flatten_output layer of the pretrained Inception-BN was stripped away and a new 3 layer fully connected neural net was added on top to convert it to a classifier of Normal vs Nodular Chest X-Ray Images.
| Property | Values |
|---|---|
| Pretrained Model | Inception BN |
| Optimizer used | stochastic gradient descent(SGD) |
| Learning rate | 0.01 (network is very sensitive, LR=.1 and it never converges) |
| Mini Batch Size | 16 |
| Epochs | trained until reach 100% on training set, used network with best validation score |
| 3 new FC Layers | 512x256x128 |
| GPU trained on | Nvidia GEFORCE GTX 960M |
Accuracy : 72.02 %
openi.nlm.nih.gov has a large base of X-Ray,MRI, CT scan images publically available.Specifically Chest Xray dataset used.
This code can be used for Deep Transfer Learning on any X-Ray dataset to train using Inception-BN as the PreTrained network. You can also use any of the models in the MXNet model library to run this code. It has been tested with Resnet-34 and resnet-50 as well.
- Get images- Goes to NLM website and recursively walks all pages downloading images and metadata, the images go in "../images_all" the metadata goes in a json file defined in settings.json_data_file.
python A1_getRawImages.py
- Now seperate and balance the dataset (there are about 2706 normal images and 211 nodule ones roughly 13 to 1). So flip and copy the nodule set so that they are roughly equivelant. Then break into train,test,val sets (.7, .15,.15). Copy them to seperate folders ../images_Train/nodule ../images_Train/nodule and likewise for test and Val
python A2_processRawImages.py
- Create RecordIO files that contain efecient, concatenated binary files consisting of all images in a particular category (train, test, val)
python A3_createRecFiles.py
-
Now the heavy lifting, take a pretrained deep conv net (Inception, Resnet etc.) and strip off the final fully connected layer to form headless CNN Run train and validation images through headless CNN and gather all the outputs, these are called CNN codes, Use CNN codes to train a brand new 3 layer fully connected neural net (converges quickly). This is necessary because mxnet has no way to freeze layers at this point, plus its much faster to train) then append that TRAINED fully connected neural net onto the headless CNN and train the complete net a little (converges quickly).
then save the best complete net (use A_utilities.epoc_end_callback_kp(...) in the fit function. Whenever accuracy exceeds previous record save the net. Finaly reload complete net and test on validation datapython A4_freeze_all_but_last_layer.py
- This is a binary classifier, there is no attempt to segment the lungs, this pretraining step may help imensily
- More data
- Deeper network (Resnet 150 maybe?)
- Localization, where nodules are located