Artificial Neural Networks and BPPC Features for Detecting COVID-19 and Severity Level

Since vaccination started, the COVID-19 scenario has improved. On the other hand, although the number of deaths has significantly dropped, the number of new cases is still a concern. Thus, patient tracking and follow-up are essential tasks, and chest X-ray examination is the first-order tool. While several studies using CXR and computing have been developed, they did not translate into clinical applications yet. One of the reasons is the computational effort required to run huge deep learning models and its high cost to be adopted in community clinics. Therefore, this work proposes a lightweight (few computational resources needed), fast (training and inference time), and reasoned solution for automatic COVID-19 detection and assessment of its severity. Our method is based on extracting features by Binary Pattern of Phase Congruency (BPPC) in segmented CXR images. Radiomic features are extracted from the segmented CXR image, and an SVM-based selection process is used to build two models of a shallow Feed-Forward network. The results surpass previous studies, with an average accuracy for COVID-19 detection of 98.71%. For images without evidence of infection but with a positive PCR test, an accuracy of 94.74% is reached. In a second task, the severity level of COVID 19 is estimated with an AUC of 98.92%. This high performance helps improve the speed and accuracy of diagnosis and severity assessment of COVID19 infection, proving to be a viable option in transitioning from a research field to a clinical environment. © 2022 IEEE.

Screening of viral pneumonia and COVID-19 in chest X-ray using classical machine learning

Governments, civil society, health professionals, and scientists have been facing a relentless fight against the pandemic of the COVID-19 disease;however, there are already about 150 million people infected worldwide and more than 3 million lives claimed, and numbers keep rising. One of the ways to combat this disease is the effective screening of infected patients. However, COVID-19 provides a similar pattern with diseases, such as pneumonia, and can misguide even very well-trained physicians. In this sense, a chest X-ray (CXR) is an effective alternative due to its low cost, accessibility, and quick response. Thus, inspired by research on the use of CXR for the diagnosis of COVID-19 pneumonia, we investigate classical machine learning methods to assist in this task. The main goal of this work is to present a robust, lightweight, and fast technique for the automatic detection of COVID-19 from CXR images. We extracted radiomic features from CXR images and trained classical machine learning models for two different classification schemes: i) COVID-19 pneumonia vs. Normal ii) COVID-19 vs. Normal vs. Viral pneumonia. Several evaluation metrics were used and comparison with many studies is presented. Our experimental results are equivalent to the state-of-the-art for both classification schemes. The solution’s high performance makes it a viable option as a computer-aided diagnostic tool, which can represent a significant gain in the speed and accuracy of the COVID-19 diagnosis. © 2021 IEEE.

Automatic Orientation Identification of Pediatric Chest X-Rays

Chest radiography (CXR) is one of the first choices in epidemiological analyses such as tuberculosis, cancer, pneumonia, and, recently, COVID-19. It provides crucial information for decision making, treatment, and monitoring the evolution of clinical cases from small to high complexity. Thus, it is a valuable source of information for the study, training, research, and development of computational support to medical diagnoses. In this work, we introduce a new method for chest X-ray adjustment to identifying and correcting radiographic images orientation. So, they can be automatically rotated to a standard position. Our proposal uses structural characteristics and statistics of pixel intensity patterns of CXR images. Divided into three steps, our method begins with the preparation of the photos, followed by a feature extraction strategy, and it ends with the X-ray image orientation identification. We use three different databases that include pediatric and adult radiographic imaging. A result showed 99.4% accuracy in the databases in our experiments. The code prepared by the authors is publicly available. © 2020 IEEE.