Three-Stage Ensemble of Image Net Pre-trained Networks for Pneumonia Detection
Focusing on detection of pneumenia disease in the Chest X-Ray images, this paper proposes a three-stage ensemble methodology utilizing multiple pre-trained Convolutional Neural Networks (CNNs). In the first-stage ensemble, k subsets of training data are firstly randomly generated, each of which is then used to retrain a pre-trained CNN to produce k CNN models for the ensemble in the first stage. In the second-stage ensemble, multiple ensemble CNN models based on multiple pre-trained CNNs are integrated to reduce variance and improve the performance of the prediction. The third-stage ensemble is based on image augmentation, i.e., the original set of images are augmented to generate a few sets of additional images, after which each set of images are input to the ensemble models from the first two stages, and the outputs based multiple sets of images are then integrated. In integrating outputs in each stage, four ensemble techniques are introduced including averaging, feed forward neural network-based, decision tree-based, and majority voting. Thorough experiments were conducted on Chest X-Ray images from a Kaggle challenge, and the results showed the effectiveness of the proposed three-stage ensemble method in detecting pneumonia disease in the images. Keywords - Ensemble, Convolutional Neural Network, Chest X-Ray, Pneumonia, Imagenet Pre-Trained Model, Pneumonia Detection.