Pulmonary Infectious Disease Detection from Chest X-ray Images and Medical Records Using Deep Learning

Abstract

Pulmonary infectious diseases are major public health problems worldwide. These diseases affect lungs and if not diagnosed properly in time, they can become fatal. Chest x-ray images are widely used for pulmonary disease detection and diagnosis. Detection of pulmonary diseases from chest x-ray images is difficult and needs experience due to similar pathological features of the diseases and sometimes misdiagnosis occurs due to this. Several researchers used deep learning and machine learning techniques to solve this problem. These studies used chest x-ray images exclusively to develop pulmonary disease detection models. But using only chest x-ray images can not necessarily lead to accurate disease detection. Medical records are required to interpret chest x-ray images in the appropriate clinical context. This work intends to develop a multi-input pulmonary infectious disease detection model using chest x-ray images and medical records. Concerning this, the medical record dataset is collected from St. Peter's specialized hospital, Alert hospital, and Yekatit 12 hospital medical college. Chest x-ray images are collected from St. Peter’s specialized hospital, Alert hospital and Kaggle repository. Data transformation, normalization, and feature selection are among the preprocessing techniques applied to medical record datasets. For the chest x-ray image data, image cleaning, augmentation and normalization preprocessing techniques are performed. Convolutional neural network for image processing and multilayer perceptron for medical record processing are applied to develop the model. Feature level concatenation is performed to join the output feature vectors from convolutional neural network and multilayer perceptron for disease detection model development. For the purpose of comparison, we also developed an image-only and medical record-only model. The image-only model gives an accuracy of 92.68%, the medical record-only model results in 98.72% accuracy, and the joint model accuracy is recorded at 99.61%. Also the joint model scored accuracy of 93.94% in medical expert evaluation.