Enhancement of Chest X-Ray Images to Improve Screening Accuracy Rate Using Iterated Function System and Multilayer Fractional-Order Machine Learning Classifier

Chia-Hung Lin,Chien-Ming Li,Neng-Sheng Pai,Pi-Yun Chen,Jian-Xing Wu,Ying-Che Kuo

doi:10.1109/jphot.2020.3013193

Chia-Hung Lin, Chien-Ming Li + Show 4 more

Open Access

https://doi.org/10.1109/jphot.2020.3013193

Copy DOI

Abstract

Chest X-ray (CXR) images are usually used to identify the causes of patients’ symptoms, including the classes of lung or heart disorders. In visualization examination, CXR imaging in anterior–posterior (A–P) views is a preliminary screening method used by clinicians or radiologists to diagnose possible lung abnormalities, such as pneumothorax (Pt), emphysema (E), infiltration (In), lung cancer (M), pneumonia (P), pulmonary fibrosis (F), and pleural effusion (Ef). However, the identification of the causes of multiple abnormalities associated with coexisting conditions presents a challenge. In ruling out a suspected lung disease, the signs and symptoms of physical conditions need to be identified to arrive at a definitive diagnosis. In addition, low contrast CXR images and manual inspection restrict automated screening applications. Hence, this study aims to propose an iterated function system (IFS) and a multilayer fractional-order machine learning classifier to rapidly screen the possible classes of lung diseases within regions of interest on CXR images and to improve screening accuracy. For digital image processes, a two-dimensional (2D) fractional-order convolution is used to enhance symptomatic features. The IFS with nonlinear interpolation functions is then used to reconstruct the 2D feature patterns. These reconstructed patterns are self-affine in the same class and thus help distinguish normal subjects from those with lung diseases. The accuracy rate is thus improved. Pooling is performed to reduce the dimensions of the feature patterns and speed up complex computations. A gray relational analysis-based classifier is used to identify the possible classes of the signs and symptoms of lung diseases. For digital CXR images in A-P view, the proposed multilayer machine learning classifier with k-fold cross-validation presents promising results in screening lung diseases and improving screening accuracy rate relative to traditional methods. The proposed classifier is evaluated in terms of recall (99.6%), precision (87.78%), accuracy (88.88%), and F1 score (0.9334).

Highlights

Lung diseases may refer to several types of diseases or disorders that affect the pulmonary functions in one or both sides of lungs and the right/left upper, middle, or lower lung regions; they are caused by influenza, infection, tuberculosis, pulmonary edema, aspiration pneumonia (AP), and lung cancer, which can lead to breathing problems or acute respiratory distress syndrome (ARDS)
270 subjects were collected from the Chest X-ray (CXR) database; the subjects included 230 individuals from the National Institutes of Health (NIH) CXR database [30] with typical lung diseases (N, Pt, E, In, M, P, F, and Ef) and 40 subjects from COVID-19 CXR database [41] with novel COVID-19 (25), severe acute respiratory syndrome (SARS) (7), Streptococcus pneumoniae pneumonia (5), and ARDS (3)
For the feature patterns fed with 2D fractional-order convolution (FOC) and iterated function system (IFS) processes, the gray relational analysis (GRA)-based classifier could improve the accuracy rate from 83.48% to 88.88% to separate “disease present” from “disease absent” and enable clinicians to respond for treatment actions

Summary

Introduction

Lung diseases may refer to several types of diseases or disorders that affect the pulmonary functions in one or both sides of lungs and the right/left upper, middle, or lower lung regions; they are caused by influenza, infection, tuberculosis, pulmonary edema, aspiration pneumonia (AP), and lung cancer, which can lead to breathing problems or acute respiratory distress syndrome (ARDS). Smoking is the most common cause of respiratory diseases and a common risk for lung cancer. The incidence of this case has increased in about 25% of lung cancers patients who are never-smokers [1], [2]. Several lung diseases, such as asthma and lung cancer, are caused by environmental factors. Subjects with asthma have improved breathing flow rates compared with those with chronic obstructive pulmonary disease (COPD). COPD, emphysema (E), and chronic bronchitis are serious respiratory diseases and associated with cigarette smoking; they reduce air flow and cause difficulty in breathing.

Objectives

Methods

Results

Conclusion