Convolutional Neural Network with Attention Modules for Pneumonia Detection

Abstract

In 2017, pneumonia was the primary diagnosis for 1.3 million visits to the Emergency Department (ED) in the United States. The mortality rate was estimated to be 5%-10% of hospitalized patients, whereas it rises to 30% for severe cases admitted to the Intensive Care Unit (ICU). Among all cases admitted to ED, 30% were misdiagnosed, and they did not suffer from pneumonia, which raises a flag for the need for more accurate diagnosis methods. Several methods for pneumonia detection were recently developed using AI in general and more specifically, using deep neural networks. Even though it worth acknowledging the significant limitations and concerns on the generalizability of such models and the barriers facing the employment of this technology for clinical practice. In this paper, an Attention model is used with a Convolutional Neural Network (CNN) for lung pneumonia diagnosis. The backbone of the model is a ResNet50 architecture with an added dual attention layer. The model was trained on the chest x-ray dataset for the aim of chest pneumonia classification. The model achieved an average validation accuracy of 97.82% and AUROC of 0.98842 on our split with cross validation. Regarding the original split, accuracy was 77.63% and AUROC of 0.7967 on the official test set. In summary, incorporation of established computer vision techniques such as Attention modules seems to be a promising approach for advancing medical image analysis.

Publication
In 2020 International Conference on Innovation and Intelligence for Informatics, Computing and Technologies (3ICT)
Ahmed H. Shahin
Ahmed H. Shahin
PhD Student at University College London

My research is focused on developing machine learning approaches for medical imaging applications.

Related