Open AccessA Novel Deep Learning Approach for Detection of Pneumonia from Chest X-Rays
Author(s) -
Andrew Yuan
Publication year - 2021
Publication title -
journal of student research
Language(s) - English
Resource type - Journals
ISSN - 2167-1907
DOI - 10.47611/jsrhs.v10i3.1627
Subject(s) - computer science , convolutional neural network , transfer of learning , artificial intelligence , artificial neural network , machine learning , binary classification , deep learning , tier 2 network , binary number , support vector machine , telecommunications , arithmetic , mathematics
Current machine learning models for pneumonia detection perform poorly compared to those for other diseases. I aim to use a novel approach to build a model that can improve on performance and be applicable to other diseases.
My approach uses multi-tier neural networks instead of using single convolutional neural networks (CNNs) like what other researchers have done. My multi-tier model consists of three tier-1 Neural Networks (NNs) and one tier-2 NN. For the tier-1 NN, I use 3 selected ImageNet-trained CNNs (ResNet152V2, DenseNet201, and NASNet Hub) as the starting bases, and train each of them via transfer learning. Then, a tier-2 NN is employed to combine the prediction results from the well-trained tier-1 NNs. The tier-2 NN model is trained with the same dataset. It produces the final predictions with substantially improved performance. The dataset used for my model is pre-processed by me, and based on a public chest X-ray dataset from the NIHCC.
My model achieved an AUC score of 76.5%, which is better than each of the tier-1 NNs alone and better than most existing models created by others.
My multi-tier approach accurately detects pneumonia from chest X-rays. It’s practical and it employs incremental learning, meaning it can be continuously improved over time. In the future, I could extend my model from binary classification to multiclass classification, apply it for lung cancer, or even other diseases.