Inverted bell-curve-based ensemble of deep learning models for detection of COVID-19 from chest X-rays

Neural Computing and Applications https://doi.org/10.1007/s00521-021-06737-6 (0123456789().,-volV)(0123456789(). ,- volV) S.I.: AI-BASED E-DIAGNOSIS Inverted bell-curve-based ensemble of deep learning models for detection of COVID-19 from chest X-rays Ashis Paul1 • Arpan Basu1 • Mufti Mahmud2,3,4 • M. Shamim Kaiser5 • Ram Sarkar1 Received: 8 March 2021 / Accepted: 21 September 2021 Ó The Author(s) 2021 Abstract Novel Coronavirus 2019 disease or COVID-19 is a viral disease caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). The use of chest X-rays (CXRs) has become an important practice to assist in the diagnosis of COVID-19 as they can be used to detect the abnormalities developed in the infected patients’ lungs. With the fast spread of the disease, many researchers across the world are striving to use several deep learning-based systems to identify the COVID-19 from such CXR images. To this end, we propose an inverted bell-curve-based ensemble of deep learning models for the detection of COVID19 from CXR images. We first use a selection of models pretrained on ImageNet dataset and use the concept of transfer learning to retrain them with CXR datasets. Then the trained models are combined with the proposed inverted bell curve weighted ensemble method, where the output of each classifier is assigned a weight, and the final prediction is done by performing a weighted average of those outputs. We evaluate the proposed method on two publicly available datasets: the COVID-19 Radiography Database and the IEEE COVID Chest X-ray Dataset. The accuracy, F1 score and the AUC ROC achieved by the proposed method are 99.66%, 99.75% and 99.99%, respectively, in the first dataset, and, 99.84%, 99.81% and 99.99%, respectively, in the other dataset. Experimental results ensure that the use of transfer learning-based models and their combination using the proposed ensemble method result in improved predictions of COVID-19 in CXRs. Keywords COVID-19 detection  Convolutional neural network  Ensemble learning  Chest X-ray  Bell-shape function 1 Introduction The Novel Coronavirus 2019 disease or COVID-19 caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) is spreading rapidly all over the globe. The & Mufti Mahmud & Ram Sarkar 1 Department of Computer Science and Engineering, Jadavpur University, Kolkata 700032, India 2 Department of Computer Science, Nottingham Trent University, Clifton, Nottingham NG11 8NS, UK 3 Medical Technologies Innovation Facility, Nottingham Trent University, Clifton, Nottingham NG11 8NS, UK 4 Computing and Informatics Research Centre, Nottingham Trent University, Clifton, Nottingham NG11 8NS, UK 5 Institute of Information Technology, Jahangirnagar University, Dhaka 1342, Bangladesh World Health Organization (WHO) declared it as a global pandemic [1] on March 11, 2020, and as of January 2021, the virus has infected more than 105,000,000 people worldwide. Though having a lower mortality rate than its predecessors, Severe Acute Respiratory Syndrome (SARS) and Middle East Respiratory Syndrome (MERS), COVID19 has killed more than 2,200,000 people worldwide. The standard and the definitive way to detect COVID-19 is via Reverse Transcription Polymerase Chain Reaction (RTPCR). However, such tests are reported to have a high falsenegative rate [2] and variable sensitivity. So as an alternative diagnosis method and to determine the progress of the disease in a patient’s body, chest X-rays (CXRs) and computed tomography (CT) scans are used [3]. This is due to the fact that COVID-19 causes visible abnormalities in the lungs which are visually similar yet often distinct from viral pneumonia [4]. Though chest CT scans have high sensitivity towards pulmonary diseases, they are not portable and carry a high risk of exposing health workers and the person under investigation to the virus. The CXRs being portable are considered to be a safe 123 Neural Computing and Applications alternative [5] as the person under investigation can be imaged in a more isolated environment, thereby lowering the risk of spreading the virus. Although a vaccine has been developed, it will take time to vaccinate the entire world population, especially in developing countries [6]. With the recent developments in data-driven Deep Learning (DL), various DL models like convolutional neural networks (CNNs) are being used extensively to study medical images [7]. CNNs are achieving state-of-the-art performances in classification into disease classes for diagnosis and also in segmentation of the region of interest (ROI) in medical images. This is enabled by the fact that CNNs can learn local features very accurately from a given medical image which can be a CT scan or a CXR. Combining outputs of multiple classifiers to generate the final output is a popular approach to enhance the performance of classification. The combination of ensemble algorithms works on the output scores of the individual classifiers, which may have different architectures to capture different elements of data or different input vectors generated from the same data instance [8]. Existing popular rank level or confidence score level ensemble methods like majority voting, sum-rule (soft voting) [9] focus on a linear combination of the classifiers’ outputs to generate the final prediction, lacking any consideration of the output vector quality. In this paper, we propose a novel weighted average ensemble method to combine the confidence scores of various pretrained CNN models to achieve better performance in detecting COVID-19 from CXR images. The inverted bell curve is used to assign weights to the classifiers’ outputs. The more we move further from the centre of the bell we attain higher weight values, and thus the shape of the inverted bell is utilized to calculate the weight for an output vector. Both the classifiers’ output quality and the overall performance of the classifiers are considered, thereby providing a more justifying combination of classifier outputs. Transfer learning is used where the CNN models are first pretrained on a huge dataset to learn basic image-related features. Then they transfer the knowledge with some fine-tuning to classify CXR images to help the medical practitioners in the diagnosis of COVID-19. We highlight the benefits of the proposed inverted bell-curvebased ensemble method to improve the accuracy and robustness of these transfer learning models. To summarize, the contributions of this work are as follows: 1. We propose an ensemble of transfer learning models to classify CXR images to detect COVID-19. 2. We propose a novel ensemble method that uses an inverted Bell curve to assign weight to the output of the classifiers and performs weighted average to obtain the final output vector. 123 3. The proposed approach is evaluated on COVID-19 Radiography Database [10] and IEEE COVID Chest X-ray Dataset [11] and state-of-the-art results are o (...truncated)