2021 International Seminar on Intelligent Technology and Its Applications (ISITIA) | 978-1-6654-2847-7/21/$31.00 ©2021 IEEE | DOI: 10.1109/ISITIA52817.2021.9502264 2021 International Seminar on Intelligent Technology and Its Applications (ISITIA) Impact of Aligning Saliency Maps on COVID-19 Disease Detection Using Chest X-Ray Images Ardimas Andi Purwita and Nunung Nurul Qomariyah Computer Science Department Faculty of Computing and Media Bina Nusantara University Jakarta, Indonesia 11480 {ardimas.purwita,nunung.qomariyah}@binus.edu Abstract—The Coronavirus disease 2019 (COVID-19) has been spread across the world in the year 2020. During the same period, many deep learning researchers have proposed different screening or diagnostic methods as an alternative to the com- monly used method, e.g., reverse-transcriptase polymerase chain reaction (RT-PCR). One of the alternatives is the use of chest X- ray (CXR) images. In this paper, we first highlight the fact that by using public, pretrained deep learning model can yield a bias result. For example, by applying a saliency map, we show that a model point to features that are located outside of the lungs. In addition, by applying multiple saliency maps, differences in locations where a model focuses on can be observed. Therefore, we propose a new loss function where we constraint the saliency maps to converge to the same region. The results show that the proposed method is better compared to the model without alignment, where the Fl-score of the proposed model is 91.3% versus 89.2%. Index Terms—deep learning, COVID-19, saliency mapsI . I. I ntroduction A novel coronavirus that was first recognized in Wuhan, China, in 2019 was referred to as coronavirus disease 2019 (COVID-19) by the World Health Organization (WHO) on 11 February 2020 [1]. The COVID-19 mainly disrupts the respiratory system that may lead to respiratory failure causing deaths to many people [2], Due to the surge of the number of cases since its first introductory in the late of 2019, the WHO declared the COVID-19 as a pandemic on 11 March 2020 [3]. As a response for the pandemic, many researchers published their works regarding the COVID-19 in 2020, for example, from the domain of machine learning (ML)-based computer vision, there are 2,212 studies that are published for the topic of diagnosis or prognosis of COVID-19 by using chest radiographs (CXR) or chest computed tomography (CT) images [4]. One of the important claims from [4] is that none of the models can be used in the real world, which is mainly due to methodological flaws or biases. Specifically, issues with biases in [4] are identified by using a tool called the prediction model risk of bias assessment tool (PROBAST), which covers four domains, i.e., participants, predictors, outcomes, and analysis [5]. A consequence of biases is shown in [6, Figure 11], where the Gradient-weighted Class Activation Mapping (Grad- CAM) [7] is used to visualize if an ML model focuses on the (a) Original Image (b) Image with Grad-CAM Fig. 1. Grad-CAM looks at region outside of the lungs. right portion of an image. It is shown in [6] that the Grad- CAM results show that a model might focus on the region that is outside of the lungs in some images. For example, by using a pre-trained model from [8] on new, unseen images and apply the Grad-CAM, Fig. 1 shows regions where the ML model looks at when it classifies the images. That is, the model mistakenly looks at the text annotations that are embedded in the images. In order to have a trustworthy ML model, it needs to focus on the lung region and not focus on auxiliary annotations [6]. In other words, an ML model should be able to show that all decisions must be due to features that are part of human anatomy. For example, a feature example that Grad-CAM should highlight is the region where ground glass opacities are located [9]. A solution to the previously mentioned problem is the use of lung segmentation before the training phase as discussed in [6] or [10]. Consequently, their ML models are constrained to learn from features that are located inside the region of interest, i.e., lungs. In this paper, we want to focus on another problem 396 © IEEE 2021. This article is free to access and download, along with rights for full text and data mining, re-use and analysis.