Overview / Usage

The risk of pneumonia is immense for many, especially in developing nations where billions face poverty and rely on polluting forms of energy. According to WHO, Pneumonia accounts for 15% of all deaths of children under 5 years old, killing 808694 children in 2017. Chest X-rays are currently the best available method for diagnosing pneumonia, playing a crucial role in clinical care. Detecting pneumonia in chest radiography can be challenging and which relies on the availability and the skill of radiologists. The appearance of pneumonia in X-ray images is often unclear and can overlap with other diagnoses. In this work, we present a model that can automatically detect pneumonia from frontal-view chest X-rays. In recent times, Convolutional Neural Networks (CNN)-motivated deep learning algorithms have become the standard choice for medical image classifications as compared to traditional computer vision techniques. We have used transfer learning while training our network which not only increases our accuracy but also decreases computational power and time. Transfer learning is useful when we are working with a dataset which is relatively small with respect to the pre-trained network we are working with. Although our base model, VGG16 have a lot of parameters to train we have made our architecture excluding the fully-connected layer of the base model and trained our model over our dense architecture. We have frozen the first layers of the base model and unfreeze the others. Jointly training both the newly-added classifier layers and the last layers of the base model allows us to “fine-tune” the higher-order feature representations in the base model in order to make them more relevant for the specific task. Automatic detection of diseases from chest X-rays would be very beneficiary for the medical world and it would also be helpful in the delivery of health care to populations with very less access to it and live far from places where there is access to diagnostic imaging specialists and expert radiologists.

Methodology / Approach

The overall architecture of the proposed CNN model consists of two major parts: the VGG16 layers and then the fully connected layer.

In our proposed model The VGG layers are used as feature extractors. They work as our base layer. VGGNet consists of 16 convolutional layers and is very appealing because of its very uniform architecture. Here we load the pre-trained network with weights trained with ImageNet [10]. However VGG16 network is a very big architecture and consists of 138 million parameters which are difficult to compute. So in our model, we load the network that doesn’t include the classification layers, which decreases the number of the parameters to 14 million, which makes the total number of layers to be 19. As you go higher up, the features are increasingly more specific to the dataset on which the model was trained. So here we unfreeze the layers after layer number 15, which makes the total number of trainable parameters to almost 7 million.

After the base model, we flatten the last layer of the base model, and then we follow with two dense networks of neuron numbers of 128 and 64 respectively. In both of the dense layer, we use Rectified Linear unit (ReLu) as our Activation function. Following the dense layers, we add a Dropout layer of the rate of 0.25. In the last layer, we add a Softmax function activated layer of size 2 which performs the classification task.