JMIR Publications

ABSTRACT

Background:

The coronavirus disease of 2019 (COVID-19) has been declared a pandemic and has raised worldwide concern. Lung inflammation and respiratory failure are commonly observed in moderate-to-severe cases. Chest X-ray (CXR) imaging is compulsory for diagnosis, and interpretation is commonly performed by skilled medical specialists. Many studies have been conducted using machine learning approaches, such as deep learning (DL), with acceptable accuracy; however, other dimensions such as computational time have been much less discussed.

Objective:

The motivation of our work is to develop a new computer-aided diagnosis (CADx) tool for identifying CXR images of COVID-19 infection using multiple machine learning techniques based on multi-layer classification architecture. It operated under the condition of minimal computational time with promising classification results.

Methods:

In this retrospective study, five public datasets of 4,200 CXR images were analyzed using multiple machine learning techniques which include decision tree (DT), support vector machine (SVM), and neural networks (NNs). First, image segmentation, image enhancement, and feature extraction techniques were performed. Second, machine learning techniques were selected based on classification performance. Finally, the selected machine learning techniques were assembled into a Multi-Layer Hybrid Classification model for COVID-19 (MLHC-COVID-19). Specifically, the MLHC-COVID-19 consists of two layers, Layer I: Healthy and Unhealthy; Layer II: COVID-19 and non-COVID-19.

Results:

The MLHC-COVID-19 is evaluated with real COVID-19 cases from various databases. The classification results show promising performance with minimum processing time achieving accuracy, sensitivity, and specificity of 0.962, 0.962, and 0.971, respectively. This demonstrates the effectiveness of the MLHC-COVID-19 in classifying CXR images, enhancing the accuracy of CXR image interpretation with a reduction in the interpretation time. A web-based MLHC-COVID-19 CADx (http://psuva.com/mlhc) has been developed for public use.

Conclusions:

MLHC-COVID-19 performed with promising classification results. A comparison between the MLHC-COVID-19 and other state-of-the-art DL techniques has been presented and discussed. MLHC-COVID-19 could be improved in future work. Some limitations are clearly stated. The entire process is semi-automated. The CXR raw images must be pre-processed before being classified with MLHC-COVID-19.