IEEE Computational Intelligence Magazine - November 2020 - 14

They collected a new fully-annotated LUS image dataset from
several hospitals in Italy, and the labels indicate the severity of
the disease at the frame, video, and pixel levels (segmentation
masks). Using these data, several artificial neural network models have been developed to solve the tasks related to the automatic analysis of LUS images. To predict the severity of the
disease associated with an input frame, an extension of spatial
transformer networks was proposed, which can provide localization of the diseased area in a weakly supervised manner. To
conduct scoring at the video level, an effective frame score
aggregation function was proposed, and three artificial neural
networks, vanilla U-Net [25], U-Net++ [26], and Deeplabv3+
[27], have been adopted for the segmentation of COVID-19
imaging biomarkers at the pixel level.
Wang et al. [15] proposed a CNN-based model, COVIDNet, for detecting COVID-19 infected patients from a dataset
of chest radiography images. The dataset is an open dataset consisting of 13,800 images collected from 13,725 patients. COVID-Net utilizes a novel lightweight residual block, the
projection-expansion-projection-extension (PEPX), to improve
representational capacity while maintaining reduced computational complexity. Furthermore, COVID-Net is designed to
make predictions using a qualitative analysis method called
GSInquire, to obtain deeper insight into crucial features related
to COVID-19 infected patients, which can assist clinicians in
efficient and precise diagnosis.
Han et al. [16] presented an attention scheme involving
deep 3D multiple instance learning called AD3D-MIL to learn
a detection model from 3D chest CTs. With the attention
scheme AD3D-MIL, not only can it accurately predict an individual category of disease such as COVID-19, common pneumonia, or no pneumonia, but it also produces interpretability
of results. During the learning process, users will not receive a
set of labeled instances, but each bag contains many instances,
TABLE I Issues addressed by existing neural
network methods.
TPVP



COVID-NET [15]



HAN ET AL.'S
STUDY [16]



PANWAR ET AL.'S
STUDY [17]



OH ET AL.'S
STUDY [18]



APOSTOLOPOULOS
ET AL.'S STUDY [19]



WANG ET AL.'S
STUDY [20]



AYYOUBZADEH
ET AL.'S STUDY [21]
VAID ET AL.'S
STUDY [22]

CSVI

ROY ET AL.'S
STUDY [14]



TRD

PHPM




IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | NOVEMBER 2020

and each bag has a label rather than separately labeled sets of
instances. The idea behind AD3D-MIL is to treat all CT images of an individual patient as the instances of a labeled bag.
Meanwhile, a fully 3D convolutional neural network is used to
produce the feature map of each instance, and an attentionbased MIL pooling is designed to select and combine the feature maps into a bag representation. Finally, the bag re---presentation
is fed into a typical fully-connected neural network to make
the final predictions.
Panwar et al. [17] developed a deep learning-based COVID-19
detection model that can detect a COVID-19 positive patient
within 5 seconds using X-ray images. The proposed model
extends VGG-16 by adding five custom layers as the head layers, of which the first layer is an average pooling 2D layer.
Unlike max pooling, this average pooling layer uses the average
value of all the pixels with a pool size of (4, 4) to down-sample
the images. The second layer is a flattened layer that transforms
a two-dimensional tensor into a vector as an input of a fully
dense connected layer (i.e., the third layer). Meanwhile, the
activation function of the fully dense connected layer is ReLU.
The fourth layer is a dropout layer that ignores half of the units
of the fully dense connected layer. The fifth layer is the output
layer, which uses two units to produce the confidence values
for the infected and uninfected, respectively. Based on a pretrained VGG-16 with the five layers added, the proposed model
was able to achieve a 97.62% true-positive rate with a limited
amount of data, consisting of 142 images of uninfected and 192
images of infected people.
The training of neural networks with limited training sample sizes is key to applying deep learning to address the issues
regarding COVID-19. To deal with the limited data size, Oh et
al. [18] developed a neural network for COVID-19 diagnosis
that is suitable for training with limited X-ray images. An
extended fully convolutional neural network called (FC)DenseNet103 [28] was adopted for lung image segmentation.
The results of the lung image segmentation from the segmentation networks are utilized for masking the pre-processed
images. To classify the masked images, ResNet-18 [29] was
adopted to build a classification model. Meanwhile, the classification model was implemented with two different contexts:
global appearance and zooming in a partial area. To consider
the view of global appearance, each masked image is resized to
224 × 224 so that each input is a complete X-ray image. Oh et
al. utilized this approach as a baseline network for experimental
evaluation. To consider zooming in a partial area of an X-ray
image, each masked image is cropped randomly to produce
several 224 × 224 images so that a masked image may produce
several input images. Although the overall accuracy of this
approach is 91.9%, slightly lower than that of COVID-Net [15]
(92.4%), the model size of this approach (11.6 M parameters) is
much smaller than that of COVID-Net (116.6 M parameters).
In other words, this approach requires much less data to train
the model.
Apostolopoulos et al. [19] examined the significance of the
extracted features and utilized MobileNet V2 [30] to train a

IEEE Computational Intelligence Magazine - November 2020

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - November 2020