Transfer learning approach for multi-disease classification using chest X-ray images

Abstract

A large portion of deaths in the world is caused by thorax diseases. They are caused by fungi, bacteria and viruses. Radiologists find it hard to identify the disease just by seeing the X-ray images as a patient may have multiple diseases which may be overlapped over other diseases. The main aim of this study is to help radiologists to detect the disease with the probability of other diseases. We proposed the architecture of a deep learning (DL) model which is used for identifying the thorax diseases using the transfer learning model which would reduce the vast time and model complexity. National Institute of Health Chest X-ray dataset is used for image pre-processing which contains more than 1 lakh images of around 30,000 unique patients with 14 different types of thorax diseases, downscaled to 256*256*3 which are further augmented and fed to different neural network models pre-trained on ImageNet dataset. We prepared three different models DenseNet121, MobileNet, and InceptionV3, and we analysed the performance. We used an ensemble model – voting classifier, for combining the output from all pre-trained models. A voting classifier model named soft voting is used which gets trained on a group of numerous models, here three models. The outputs of all neural networks which are pre-trained are summed into a prediction vector by taking the mean of probabilities of all the three models and it then outputs the majority of three diseases probabilities into the final prediction vector associated with that X-ray image.