Detection of COVID-19 Lung Lesions in Computed Tomography Images Using Deep Learning

The novel coronavirus (COVID-19) is a disease that mainly affects the lung tissue. The detection of lesions caused by this disease can help to provide an adequate treatment and monitoring its evolution. This research focuses on the binary classification of lung lesions caused by COVID-19 in images of computed tomography (CT) using deep learning. The database used in the experiments comes from two independent repositories, which contains tomographic scans of patients with a positive diagnosis of COVID-19. The output layers of four pre-trained convolutional networks were adapted to the proposed task and re-trained using the fine-tuning technique. The models were validated with test images from the two database’s repositories. The model VGG19, considering one of the repositories, showed the best performance with 88% and 90.2% of accuracy and recall, respectively. The model combination using the soft voting technique presented the highest accuracy (84.4%), with a recall of 94.4% employing the data from the other repository. The area under the receiver operating characteristic curve was 0.92 at best. The proposed method based on deep learning represents a valuable tool to automatically classify COVID-19 lesions on CT images and could also be used to assess the extent of lung infection.


INTRODUCTION
Coronavirus disease 2019  is caused by the severe acute respiratory syndrome coronavirus type 2 (SARS-CoV-2). It primarily affects the human respiratory system and represents the seventh member of the coronavirus family that infects humans [1] . The first case, identified as viral pneumonia until then, appeared in late December 2019 in Wuhan, China.
According to the records issued by the World Health Organization, until September 1st, 3,341,264 cases have been registered in Mexico and 217,558,771 around the world [1] [2] [3] . It is known that the COVID-19 infection has an incubation period from 1 to 14 days, which varies depending on some human characteristics like the status of the immune system and the age [1] . In Mexico, coronavirus cases are classified by stages according to their severity, clinical stage and signs presented: stage 1 (early infection), stage 2 (pulmonary stage) and stage 3 (hyperinflammatory stage) [4] .
The reverse transcription -polymerase chain reaction (RT-PCR) tests represent the main method to detect COVID-19, providing results with a specificity close to 100% [5] ; however, when using this standard test as a reference, some drawbacks must be considered. For example, a low sensitivity (59% -79%) has been observed during the early phase of the disease [5] [6] [7] . Due to the continuous evolution and genetic diversity that the new coronavirus has presented, the results of clinical tests can be affected by the variation in the viral ribonucleic acid (RNA) sequence [8] . Also, it is import to remark that the diagnostic period can vary from 5 to 72 hours [9] .
The study presented by Uysal et al. [10] found that 25% of asymptomatic patients, diagnosed with an RT-PCR test, did not show signs of lesions on their computed tomography (CT) scans, while the rest showed abnormal findings associated with lesions similar to those in patients with symptoms. The most common signs were ground glass opacity (GGO), pure or with consolidation or crazy-paving patterns. Thus, due those find-ings some authors emphasize over the importance of performing RT-PCR tests in conjunction with imaging procedures such as CT to increase the accuracy of the diagnosis, injury identification, and in this way provide an adequate patient management [11] .
To confirm the coronavirus disease, the chest CT in conjunction with clinical manifestations and the epidemiological evidence have become a fundamental diagnostic tool. However, discrepancies have been reported between the results of laboratory tests and the characteristics observed in diagnostic images [12] .
Recently, some studies have shown that the CT scan of patients (asymptomatic or those in whom the result of RT-PCR test was negative) depicts abnormal signs that can be useful for the disease detection, where these studies have reported a sensitivity between 88% and 98% [5] [6] [13] [14] . The advantage of CT diagnosis lies in its short exploration time and the high resolution of the acquired image, useful for detecting and classifying lung lesions.
At present, most of the expert researchers in the clinical applications of Artificial Intelligence (AI) have focused on the diagnosis of patients with COVID-19 through the processing of medical images, addressing the analysis of findings observed in chest x-rays and/ or CT scans [15] [16] . There are several approaches that aim to take advantage of machine learning (ML), especially deep learning, to diagnose CT scans using binary pathway convolutional neural networks (CNN) (positive vs. negative) or multiple classification (healthy vs. COVID-19 versus other types of pneumonia) [16] . An example of this is the COVNet architecture performed by Li et al., which classifies positive results for COVID-19, community acquired pneumonia or negative for any lung disease through a three-dimensional CNN constituted by the ResNet50 architecture, resulting in 90% of sensitivity and a specificity of 96% [17] . Similarly, Yang et al. in [13] , built a publicly available database of CT scans of COVID-19 patients that could be used to train deep learning models. This database was subsequently used to develop an algorithm to classify COVID-19 patients in a binary way, obtaining an accuracy of 83% and an area under the receiver operating characteristic curve (AUC-ROC) of 0.95. Other work that uses deep learning techniques developed a model called CTnet-10 obtaining an accuracy of 82.1%. The authors also tested models such as DenseNet169, VGG16, ResNet50, InceptionV3 and VGG19, obtaining an accuracy of 94.52% with the latest network [18] . On the other hand, in [19] the authors attempted to segment lung lesions associated with COVID-19, reaching specificity values of up to 100% in specific tasks and models tested, but with a very low sensitivity (between 1.2% and 64.8%).
As mentioned before, a large percentage of asymptomatic patients already have abnormal findings on their CT scan images whose lesion patterns are similar to those found in symptomatic patients. In this sense, it is very important to detect these patterns in CT images to allow physicians to know if a patient has lung lesions and thus guide their treatment.
The purpose of this investigation is to detect the presence or absence (i.e., a binary classification) of lung lesions due to COVID-19 in images originated from chest CT studies using deep learning. It could be useful when it is desired to identify whether the lesions are disseminated in a large part of the lung tissue, indicating that the lesions occur in many slices of the CT study; this detection can even be valuable in assessing the evolution of lung tissue damage, and thus provide adequate treatment to the patients.

MATERIALS AND METHODS
The database used in this research corresponds to "COVID-19 CT Lung and Infection Segmentation Dataset" [20] . The images are in NIfTI (Neuroimaging Informatics Technology Initiative) format and were  where GGO is indicated with green arrows, consolidation is surrounded by segmented red ovals, pleural effusion is pointed with a yellow arrow, and crazy-paving pattern is enclosed by a blue line (also indicated by the blue arrow). In Figure 1, the images (a, b, c) belong to the Coronacases Initiative repository, and the images (d, e) correspond to the Radiopaedia repository. There are images that present inconspicuous abnormalities that could be challenging for both an inexperienced radiologist and an automatic detection model.
For instance, Figure 2 shows an example of two images from a CT scan of the same patient. The slice in a) shows slight evidence of GGO, while in b) no abnormalities or lesions are observed. Thus, the detection system must be able to identify negligible lesions that commonly appear at the early stage of the disease.  Table 1 shows the division of the data set into subsets: training, validation, and testing.

TABLE 1. Number of images for the training, validation, and test subsets (Coronacases + Radiopaedia repositories).
The CT volumes belonging to the Radiopaedia database were previously pre-processed with a pulmonary window [-1250, 250] [19] . The image format was con-

Implementation of convolutional neural networks
The algorithm was developed in Python. The implementation of the network models was carried out by means of transfer learning and subsequent fine-tuning. Transfer learning is a technique that takes advantage of existing knowledge to solve problems from a source domain to a destination domain in which, although the same task is not performed, both tasks have a certain similarity. Thus, the purpose is to solve a learning problem using the knowledge acquired by solving similar tasks [21] . On the other hand, the fine-tuning process applied in the context of deep  provided accuracy values between 89% and 99% [22] [23] [24] . with an accuracy of 93% [25] .
In this work, the transfer learning technique was implemented using four pretrained models belonging to the following networks: ResNet50 (RN50) [26] , VGG16 [27] , InceptionResNetV2 (IRNV2) [28] , and NASNetLarge (NNL) [29] . These networks were chosen due the well performance in large scale image recognition tasks,  [30] . For each of these networks, they employed weights obtained from training using data from the ImageNet repository [31] . ImageNet corresponds to a dataset widely used for object recognition purposes. Figure 3 shows the general configura-

Creation and evaluation of models with separate data repositories
As an additional experimentation, the division of the data by sources (Coronacases Initiative and Radiopaedia repositories) was proposed in order to assess the performance when evaluating the models with the data from the repositories separately. For these tests, both repositories were inspected with the intention of finding and removing low-quality images.
In this process, 77 slices with high opacity were   with fine-tune and the SGD optimizer were used.  lesions is observed, with an accuracy equal to or greater than 78% in four of the nine models evaluated.

RESULTS AND DISCUSSION
It is also important to mention that only one of the models presents a SP greater than 80%, which indicates that in most models, there is a tendency to misclassify the negative class (images without lesions). Figure 4 shows the accuracy performance of the nine models, which were trained with the training set (that includes data from both repositories) but evaluated with the test sets of each repository independently.
For purposes of better identification, data from the Coronacases Initiative repository is named as DB1 and data from Radiopaedia named as DB2. As it is observed in Figure 4, the accuracy in the classification of the DB1 images was superior in six of the nine models evaluated. It can be observed from Figure 4 that the models do not show a consistent fit to the data from both repositories separately; this may be due to the lung window preprocessing previously applied to the images of the Radiopaedia repository.  It is important to mention that the studied networks presented a high classification error or misclassification in slices that were located at the beginning or at the end of the scans (at the cephalocaudal ends). It could be due to the fact that these images show a reduced area of lung tissue while the rest of the tissue can generate structures similar to abnormality findings suggesting a pulmonary lesion. An example of this issue can be seen in Figure 5.   We must emphasize that in the present research, all the CT scans of the database included patients with a positive diagnosis of COVID-19, so the abnormality patterns found in the images are assumed to be indicative of lesions due to this disease. This represents a limitation in the present study since certainly other lung diseases such as interstitial pneumonia, sarcoidosis, alveolar proteinosis, carcinoma, etc., can produce similar patterns in CT scans to those found in patients with COVID-19 [10] [34] . Therefore, as future work, it is necessary to advance in this research to include patients with different lung diseases and classify the lesions according to their pathology of origin.