Disease Prediction Model Using Deep Transfer Learning

Pneumonia the word which is taken from the Greek word pneuma, breath is an infection that can be fatal as well as an inflammation of the lower respiratory tract such as the bronchioles and alveoli. It is typically brought on by inhaled bacteria as well as viruses such as Streptococcus pneumoniae, also known as pneumococcus, which has both of these characteristics. Machine learning technique called transfer learning is, where a model is applied to newly collected data instead of being learned from scratch using a convolutional neural network (CNN) model which is pre-trained that was trained on a dataset. The disease prediction model is built using transfer learning techniques. Image pre-processing methods such as image dilation, image erosion, Gaussian blur, grayscale and canny edge detection and they are carried out for extracting the characteristics of the CXR images containing 5826 image dataset which were taken from the Kaggle repository consisting of two subfolders normal and pneumonia. They were split into training, testing and validation dataset. The features are extracted from the CNN model which is built from the beginning after that transfer learning techniques are used for prediction of presence of pneumonia using the test set and in last and final step results were produced. The following three models such as CNN model got an accuracy of 78% and transfer learning model VGG16 gave an accuracy of 88% followed by the ResNet50V2 model which gave an accuracy of 90%. In all three models, ResNet50V2 model gave the highest accuracy.