Using a convulational neural network to classify CT scan images of idiopathic pulmonary fibrosis
Idiopathic pulmonary fibrosis (IPF) is the most common form of interstitial lung diseases. It affects approx. 100,000–200,000 people in the U.S. and approx. 30,000 to 40,000 new cases are seen in the U.S. every year. The clinical presentation is often as an elderly man who presents with shortness of breath and low oxygen saturation in the blood (hypoxemia). Physical examination often reveals fine crackling sounds on lung auscultation, also known as ‘Velcro’ crackles. CT scan of the lungs is characterized by subpleural cysts, traction bronchiectasis and honey-combing appearance of the lung, especially in the peripheral regions. The treatment options are limited despite recent drugs like pirfenidone, and the life expectancy is often limited to less than 5 years after the diagnosis.
In this pilot experiment, our aim was to use a convulational neural network to classify CT scan images into IPF or normal. Since, this was a pilot experiment and this is not a very common disease, our image dataset consisted of fourteen CT scan images of IPF from different patients (with the characteristic features of IPF), and nine CT images of normal lung from different patients. CT scan images of IPF patients were annotated to highlight the abnormal areas with subpleural cysts, honey-combing, etc. Google Colab Python 3.6 environment with a Nvidia GPU and 64GB RAM was used.
We used Pytorch’s torch library to split the dataset randomly into train and test datasets in 80/20 ratio. Torch vision library was used to transform and preprocess the images.
We used ResNet-50 convulational neural network (CNN) in this experiment for deep learning for image recognition and classification. The model layers are as shown below. We used optimizer= ‘Adam’.
The ResNet-50 model was trained on the train image dataset and validated on the test image dataset, starting initially with 3 epochs which were increased to achieve the threshold of maximal test accuracy and test loss above which the test loss started to increase due to overfitting of the model. The ideal number of epochs found on tuning the model parameters was seven.
The test accuracy of 100% was achieved at epoch 7 as shown in the figure above. Above epoch 7, the test accuracy was found to decrease.
Test losses after each epoch were plotted using matplotlib library and are shown in the figure below.
In conclusion, we were able to achieve 100% test accuracy in identifying CT scan images from IPF patients from normal lung CT images in this pilot study using ResNet-50 convulational neural network. We are collecting CT scan data from more IPF patients in the next stage of this project. We have also collected histopathology data from lung biopsy of IPF patients and plan to use the same approach using CNNs for image classification.
