Lung Ultrasound for COVID detection

Pablo Laso-Mielgo

, A. Ramos-Uparela

Dr. Rafael Garcia Carretero, MD, PhD

, Dr.

Angel Bueno, MD, PhD

Biomedical Engineering degree, Alcorc

on Campus, Universidad Rey Juan Carlos.

Escuela T

ecnica Superior de Ingenier

ıa en Telecomunicaci

on, Universidad Rey Juan Carlos.

Area hospitalaria, Hospital Universitario Fundaci

on Alcorc

on.

Abstract

From the Machine Learning community, US (ultrasound)

has not gained as much attention as X-Ray or CT in the

context of COVID-19. But many voices from the medical

community have advocated for a more prominent role of

US in the current pandemic [1]. Moreover, US is a non-

ionizing technique, which is increasingly gaining impor-

tance as a Point-of-Care technique.

Since two students work on this project, at most points in-

dependently, we developed different tools for COVID-19

classiﬁcation in both US videos and images. Videos may

seem more difﬁcult to deal with than images. However,

the carry more information that can be used for more ef-

ﬁcient classiﬁcation tools. Additionally, the doctor may

provide an image for COVID-19 detection. In this case,

two different tools have been developed. One is based

on pleural line detection, whereas the other takes strong,

pretrained convolutional neural networks, such as Incep-

tionV3 or VGG16. Our model is validated on an external

dataset (POCUS).

1 Introduction

This is collaborative project between URJC (Universidad

Rey Juan Carlos) and HUFA (Hospital Universitario Fun-

daci

on Alcorc

on). It has been mainly carried out by URJC

students Pablo and Alejandro, with the help of Dr. Rafael

and Dr.

Angel.

Pablo ﬁrstly focused on video preprocessing (HUFA, But-

terﬂy and POCUS dataset) and later joined Alejandro on

image classiﬁcation, which was carried out with several

different approaches. Video processing results yield a high

accuracy and generalizing ability to new data. On the

other hand, image classiﬁcation required more time, since

it was based on video frames, so most are very similar and,

even if the accuracy was very high, the generalization to

new datasets was slightly lower. However, a large pub-

lic dataset (POCUS), Transfer Learning, and a thoughtful

preprocessing effort allowed us to overcome this challenge

and achieve better results.

Transfer Learning performed well on other respiratory

pathologies such as pneumonia. For COVID-19 images,

however, it did not yield as ﬁne results as video classiﬁca-

tion, so we also created our own CNN. It is worth mention-

ing the development of a code able to detect pleural lines.

The goal of this project is processing lung images acquired

via US (ultrasound). These images are anatomically poor

and, therefore, sometimes difﬁcult to interpret. We can

also distinguish between different grades of severity, which

are scored in a range of 1-3. although no labels were given

for this task.

This project has been motivated by the need of an algo-

rithm that aids physicians to more accurately diagnose dis-

eases, especially COVID-19 [2]. Another major source of

motivation is the so-called point-of-care US, that is, pro-

vide treatment at the time and place of patient care. There

are already portable US equipment. Furthermore, using

US as a standard imaging technique (and even rather the

so typical stethoscope [3]) would be desirable. Not only

would it be faster than other imaging techniques like MRI,

but also prevent the patient from receiving high doses of

ionizing energy that other techniques, such us CT or PET,

produce.

Hence, having an algorithm that helps the medical staff

to quickly analyze and score the pathology automatically

would be ideal. It would also promote both point-of-care

US as a standard diagnosis tool, as well as foster a ben-

eﬁcial transition from ionizing imaging techniques to US,

without loss of accuracy or healthcare quality.

2 Problem Statement

This project was initially proposed by Dr.

Angel Bueno

(HUFA) as an application. It should quickly identity and

score the severity of COVID-19 in different regions of the

lungs. The images are acquired by US and are anatomi-

cally poor, so preprocessing is probably mandatory.

Lung US (LUS) is used to this aim, which is deemed to

be the most effective non-ionizing image modality for this,

and also present an alternative [4], or even overcome, CT

sensitivity for similar pathologies like pneumonia [5] [6]

[7], but also for COVID-19 [8] [9] [10] [11]. As we can

observe in Figure 1, COVID-19 patients present respira-

tory problems, so LUS can be intelligently used to identify

them.

Figure 1. Lung US Symptoms

Furthermore, as we can observe in Figure 2, the main char-

acteristics that COVID-19 patients present in LUS images

or videos are B-lines and an irregular pleural line, followed

by consolidations. A-lines, on the other hand, are idiosyn-

cratic of healthy patients.

All in all, detecting B-lines, and even an irregular pleural

line, in a patient is a very solid evidence of COVID-19.

Thus, using LUS to address the problem stated in this sec-

tion is a very promising technique.

Figure 2. Lung US Pathologies

3 Methods

In this Section we describe the procedures followed along

our project. These include also attempts that proved to be

fruitful, yet not as promising as the last results, described

in the next Section of this paper. Overall, we performed

several prepossessing steps before applying Deep Learn-

ing techniques, which have been considered to be the most

effective for this kind of tasks on previous (mostly CT)

studies [12] [13] [14].

3.1 Data Acquisition

For the development of this project, specially for the part of

training a CNN, the need of a large and labeled database is

required. The HUFA kindly helped us in this task, provid-

ing us with several images and lung US videos. We were

also instructed into the different scores that these disease

could manifest [15]. Since the need of medical images and

a larger dataset was of paramount importance for the de-

velopment of the project, we resorted to online accessible

databases on this pathology .

Later on, these videos would be divided into frames. Thus,

we could deal with images and apply a CNN (based on

convolution operations performed upon input images) to

our data.

3.2 Data Augmentation

Increasing the database using Data Augmentation was im-

plemented. This can be useful for increasing the number

of samples, or for preventing overﬁtting, by applying ran-

dom variations to the data. ImageDataGenerator [16]

allowed us to readily create new data in batches, which

also diminished computational burden. Thus, albeit low-

ering accuracy, we managed to reduce overﬁtting by creat-

ing new data from the original images, mainly by rotation,

shifting and zooming.

3.3 POCUS database

As aforementioned, the lack of images to form a quantita-

tively representative dataset was a major pitfall. To over-

come this difﬁculty that hindered the appropriate train-

ing of a CNN, we used the COVID-19 image dataset

”POCUS” [1], readily accessible on the Internet.

POCUS database includes both images and videos la-

belled as regular (healthy), COVID-19, or pneumonia. The

dataset (Figure 3) was composed of convex and linear US

probes, as well as images and video ﬁles. Images of each

class are respectively shown in Figure 4.

Figure 3. POCUS dataset size

Figure 4. POCUS image data types

3.4 Image preprocessing

The main objective of preprocessing our data set was to get

much easier for the CNN to better distinguish the proper

characteristics of the different pathologies, thus leading

with a better performance in terms of classiﬁcation. The

Figure 5. Original image

Figure 6. Original histogram

process of image preprocessing followed the following

steps:

1. Image histogram equalization

A good approach for low contrast images, character-

ized by a very narrow histogram distribution of gray

levels, is to perform a histogram equalization. This

technique allows to change the histogram grey values

into a more variable distribution, thus giving a higher

contrast image.

This process takes into account the frequency of each

grey value and transform it into a probability distri-

bution.The probability of having a certain level of

gray,denoted as p

(i), is the ratio of the number of

pixels with the value of gray i with respect to the total

pixels of the image.Continued by the calculation of

the cumulative distribution function (CDF), deﬁned

C(i) =

j=0

(j) (1)

Where 0 ¡= i ¡= L 1, and L is the total number of

possible gray levels in the image.

Finally, to get the new assigned value of each grey

value h(u):

h(u) = round



C(u) − C

min

1 − C

min



(2)

Figure above show the histogram and re-arrangement

of the resulting transformation (Figures 6 and 7) and

the original image (Figures 8 and 5).

This technique facilitate much more the posterior

techniques.

2. Pleural Line estimation

The removal of skin, soft tissue and other elements

not involving the lungs(background), is important to

facilitate to the CNN a proper pathology(foreground)

identiﬁcation in thorax ultrasound images.

Figure 7. Equalized image

Figure 8. Equalized histogram

To remove skin and soft tissue, it is necessary to ﬁrst

identify the pleural line, and for that purpose was fol-

lowing a series of steps:

• Perform a global binarization for each image

with the help of the skimage library follow-

ing a certain threshold. This threshold T

are

determined by minimizing intra-class intensity

variance, or equivalently, by maximizing inter-

class variance, determined by the ratio:

Q(s) =

inbetween

within

(3)

Each image (I

) was binarized with this thresh-

old, converting each pixel into an on-pixel or an

off-pixel.

(

0 if I

< T

1 if I

≥ T

(4)

Where I

is the brightness of the pixel.

The resulting matrix B

mx1

represents the bina-

rized image on which all future transformations

are performed.

• Afterwards, analysis units are set within the bi-

narized image. These are rectangular regions

with a height equal to the original image and

width of one pixel.

mxn

= ∪

i=1

i(mx1)

.... (5)

The total number of on-pixels were calculated

in each analysis unit B

i(mx1)

, and use it just for

knowing where the middle pointM

in which the

number of on-pixels are equally distributed is,

and storing it in a variable X (Figure 8)

X =

n











, ......,



o

(6)

Figure 9. Thresholding and X points in blue

Figure 10. Regression line of X points

With these data, we obtained a polygonal proﬁle

of the pleural membrane. Because the pleural

membrane is usually linear, we made an adjust-

ment of these points to ﬁt a line by minimiza-

tion of the squared errors. The points of the ﬁt-

ted line are stored in the variable X’, where the

trace X’is the ﬁrst approximation of the pleural

line (Figure 9)

X =

n











, ......,



o

(7)

3. Resection

After computing the approximation of the pleural

line, we proceed to extract the background informa-

tion of the original image, getting a resulting im-

age(Figure 13) without the undesired background and

perhaps make the feature extraction of the CNN eas-

ier.

3.5 Video preprocessing

Apart from preprocessing video frames as independent im-

ages, we decided to analyze a video as a whole. This would

greatly increase the generalization ability of our classiﬁer,

Figure 11. Final output before resection

Figure 12. M-mode US example

Figure 13. M-mode images VGG preprocessing

which was the main challenge we had to cope with.

For such purpose, we created a single image per each inde-

pendent video in the dataset, by means of converting each

US video into an M-mode (Motion mode) US image. As

we can see in Figure 12 at the right, a sampling line is

taken in the coordinates speciﬁed when running the code.

Note that we can assess a speciﬁc anatomical area by po-

sitioning the coordinates in the area of interest. With a

sampling line width of 5 pixels, we would take such line

from every frame of each video in the dataset. Then, ev-

ery line would be successively pasted as the column of a

new numpy array, giving rise to a new image, as can be ob-

served at the left of Figure 12. Note that long videos will

result in wider images, since there exists more frames and,

therefore, more columns that the new image will have.

After that, the resulting images would undergo standard

preprocessing, including normalization, re-sizing and his-

togram equalization. Thereafter, VGG preprocessing was

used upon them, subsequently obtaining images as shown

in Figure 13, which represents one of he batches used for

the CNN training.

4 Experiments and Results

This Section includes the results on images (after pleural

preprocessing and Transfer Learning) as well as videos.

4.1 Image Classiﬁcation

As an starting point, without any kind of processing, we

just apply some normalization to the images and input it

to an standard net, in order to know how the convolutions

layers were going to perform.

Several CNN algorithms were designed in an attempt to

correctly classify the images. Despite very bad results in

the ﬁrst CNN designs, we eventually came up with a sim-

ple design that reached an accuracy of 1.00 in the training

set as well as in the validation set, after 10 epochs. The

validation set was composed by 57 images (frames) out

of 358 altogether, all of them correctly classiﬁed into its

corresponding score. This result, albeit seemly promising,

was however not deﬁnitive. Since the images came from

three videos (each of them with a different possible score

(label)), the images were probably very similar. Further-

more, there existed the possibility that the CNN based its

decisions on characteristics of that speciﬁc video, such as

lightning or intensity, rather than the pathological signs.

In order to be able to contrast these ﬁrst results we could

not but highlight the need for a larger dataset of images,

ideally granted by the HUFA; alternatively, obtained from

public, online databases on the internet [1]. This dataset,

summed up with a the video frames from the correspond-

ing POCUS videos, conform the new dataset used for pre-

dictive purposes, with a whole new variability and quantity

of acquisitions. The validation set was composed by 653

images out of 3262 images, all of them correctly labelled.

Several architectures were tested, until we designed a sim-

ple net (Figure 14) that could give us a clue of how well the

data was organized, leading with the clue of possible noisy

information in a part of them, the algorithm was being able

to highly classify the images, but not to well generalize into

a whole new set of unknown images. The results, shown

in Figure 15, are very optimistic. We must, however, keep

in mind that these results are a consequence of overﬁtting,

so more complex models and further processing is needed.

Figure 14. First CNN

4.1.1 Transfer Learning

Instead of starting from scratch, Transfer Learning enables

us to use already trained models, albeit they were trained

for a different problem. However, we might expect it to be-

have similarly based on the problem they were originally

trained to solve. Transfer learning has the beneﬁt of de-

creasing the training time for a neural network model and

can result in lower generalization error.

Figure 15. Model accuracy

Also, we should bear in mind that (probably) the more

similar the original problem (for which the CNN was ini-

tially designed) and our challenge are, the more layers

(deeper) we will be able to maintain. This is inferred from

the fact that the ﬁrst layers will deal with low-level fea-

tures, whereas those in the middle shall be much more ab-

stract and complex, ﬁtting to the speciﬁc task, problem and

dataset they were marginally designed for. Changing just

some of these last layers might enough to train an efﬁcient

CNN.

Several models from previous studies [17] [18] were

tested, among which ’vgg base’, ’vgg cam’, ’mo-

bilenet v2’, ’nasnet’, ’dense’ and ’resnet50’. All of them

had a great tendency to classify images as COVID-19, es-

pecially resnet50. The best results were acquired with the

following models:

• VGG16

The VGG16 model was developed by the Visual

Graphics Group (VGG) at Oxford and was described

in the 2014 paper titled “Very Deep Convolutional

Networks for Large-Scale Image Recognition”. Its

structure is described in Figure 16. By default, the

model expects color input images to be re-scaled

to the size of 224×224 squares. It can be loaded

from keras.applications.vgg16 with

import VGG16 [16]. From here on, several

parameters may be re-trained. Thus, we were able

to get an accuracy of 0.69 in the validation set.

However, we noted, by plotting the confusion matrix,

that some errors were regular cases incorrectly

classiﬁed as COVID-19. However, both COVID-19

and pneumonia cases were correctly classiﬁed.

• InceptionV3

InceptionV3 is a convolutional neural network for as-

sisting in image analysis and object detection, and got

its start as a module for Googlenet. Its architecture is

shown as a simpliﬁed image in Figure 17. It is there-

fore more likely to yield relevant answers to our spe-

ciﬁc problem. Furthermore its has already been im-

plemented in a myriad of cases, some of them yield-

ing expert-level diagnostic accuracy and been proven

more efﬁcient than other previous techniques in de-

Figure 16. VGG16 structure

tecting B-lines, merged B-lines, lack of lung sliding,

consolidation and pleural effusion [19]. In this case,

the accuracy for the three is slightly lower (0.58), so

we can assert that VGG performs better. Furthermore,

just like the former VGG model, choosing more than

two epochs affect negatively the performance of the

model.

Figure 17. InceptionV3 structure

For both cases we constructed a head of a new model that

would be placed on top of the base model. Thus, we train

the last layers such that it could adapt to our speciﬁc case,

while also using a pretrained and effective model, which

led to a smaller computational burden, overall.

Both models ﬂagrantly showed an unwanted tendency

from the ”regular” to the ”COVID” class, as shown in Fig-

ures 18 and 19. However, these results, albeit unwanted,

are not so bad among all the possible incorrect answers we

could have encountered. This is because we would rather

have a normal patient diagnosed as COVID-19 than the op-

posite, that is, it has a higher cost. Armed with this reason-

ing, we should look at the Recall (TP/(TP+FN)) measured

for ”COVID” and ”pneumonia” in Tables 1 and 2 is, since

the cost of a False Negative is higher (we want to avoid at

all costs having infected people (Actual Positives) in con-

tact with others). Contrarily, we should look closer at the

Precision (TP/(TP+FP) measured for ”normal”, since the

cost of a False Positive can be, just like before, undesirable

for society. We can also use the F1-score (which takes into

account both precision and recall) for the ”normal” class.

Hence, we can readily observe that VGG performs some-

what better than Inception in this ﬁrst attempt. However,

the results are very humble. Since adding epochs do not

contribute to a better performance, we conclude from these

results that further preprocessing of our images is needed,

as well as trying other layers to re-train, or choosing an-

other head to add at the end of the architecture to better

adapt to our task.

Figure 18. VGG16 Confusion Matrix

VGG16 metrics

class Precision Recall F1

COVID-19 0.65 0.77 0.70

Pneumonia 0.83 0.91 0.87

Regular 0.74 0.49 0.59

Table 1. VGG16 metrics

Figure 19. InceptionV3 Confusion Matrix

InceptionV3 metrics

class Precision Recall F1

COVID-19 0.51 0.73 0.60

Pneumonia 0.71 0.53 0.61

Regular 0.61 0.42 0.50

Table 2. InceptionV3 metrics

4.1.2 Complex CNN

After performing a much more complex pre-processing

stage (described in previous sections) than initially, we de-

cided to use the same CNN architecture(See Figure 15)

to compare results. This time , however, we had much

more images and a strong pre-processing that showed sig-

niﬁcantly more promising results, since the generalization

adaptability was now stronger as we get rid of noisy infor-

mation. Nevertheless,the image dataset has been reduced

by the little computational capacity with which we had,

seeing the set reduced by 50 images per category. Anyway,

to try to appease this reduction, we apply a Data augmenta-

tion method in order to have a more robust set (Procedure

showed in Figure 20)

Figure 20. Data augmentation

Figure 21. Loss/Accuracy Final CNN

As shown in Figure 21 , the capability of the model gener-

alization of new incoming data has been improved.

4.2 Video Classiﬁcation

Finally, after video preprocessing had been carried out, im-

ages had multiple sizes. Therefore, they were resized to the

mean height and width, that is, (723, 512). We decided to

use the mean rather than the maximum, as usually is done,

because images were of all kind of sizes. This is because

their size depends on the video duration. We consider it

simpler to just use the mean, so that the least interpolation

possible was performed overall.

A very simple CNN was able to classify these images (one

per video) with an accuracy greater than 0.9. As schemat-

ically shown in Figure 22, the CNN consisted on two ini-

tial kernel layers. The ﬁrst layer consisting of 32 ﬁlters,

whereas the second one, 64 feature maps. All kernels had

a size of 3x3, a stride of 2, and were activated or not ac-

cording to a RELU activation function. These kernel layers

were followed by 2x2 maxpooling layers, so the output of

each was twice smaller. Thereafter, a dense layer was de-

signed along with a softmax which was used at the end

of it, so that we could convert the ﬁnal output value into

probabilities, for labelling.

We used 75% for the training set, 15% for validation, and

the rest (15%) for testing. 72 independent videos and from

different sources were used as dataset in the last attempt.

With random sampling and a batch size of 10, the folder

for the two classes (COVID-19 and healthy) were created.

By random sampling, we also doubled the initial size per

set. We chose categorical crossentropy for loss evaluation,

and accuracy as the metric to evaluate performance. Addi-

tionally, a learning rate of 0.0001 proved to be the best op-

tion, since it was not too small to get stuck in a local min-

imum, but small enough to converge towards a very good

result. Since the learning rate was small and the videos in

the dataset were of all kinds and different and utterly inde-

pendent sources, the CNN would take 8 epochs to ﬁnally

reach a 90% of validation accuracy. Henceforward, the

training accuracy improved constantly, but not so the vali-

dation accuracy, which started ﬂuctuating. Thus, we deem

it to be convenient to perform early stopping at epoch 8.

Otherwise, we would risk over-ﬁtting our model.

Figure 22. video CNN architecture

Finally, we used the testing set (20 images) to ratify our

previous assumptions. The results are shown in Figure 23

as a confusion matrix. The results reﬂected the same as the

validation set. We can observe that all COVID-19 cases

are classiﬁed as such. This is crucial, since we do not want

to have an unknown and unidentiﬁed COVID-19 vector,

which can spread the disease and lead to more patients.

On the other hand, almost all normal (healthy) cases were

classiﬁed as non-COVID19.

Taking everything into consideration, we can claim that,

even if the videos were from three different independent

sources (increasing variability), a thoughtful and complex

preprocessing of the videos (turning them into an image

per video) allowed for a very simple CNN to achieve a 90%

accuracy in both the validation and training set, added to a

great capability of generalization (preventing overﬁtting).

Figure 23. Confusion Matrix for video classiﬁcation

5 Conclusion

After several different attempts for image classiﬁcation, we

conclude that our preprocessing methods allows both for

a higher accuracy, as well as greater generalizing ability,

so that a simple CNN can readily identify COVID-19 im-

ages. Similarly, Transfer Learning, speciﬁcally VGG16,

has also proven to be effective for classiﬁcation of pneu-

monia, COVID-19 and healthy patients, with an accuracy

of 0.73%. Interestingly, it is a fairly effective option for

pneumonia detection (0.83 precision, 0.91 recall, 0.87 F1;

for pneumonia only), similar to previous studies (0.83 ac-

curacy) [20]. However, in the case of COVID-19, which is

the main goal of this paper, has worse results and, there-

fore, we discarded it in search of more efﬁcient options for

COVID-19 detection.

Video processing has rendered the most promising results

(90% validation accuracy). A very simple and basic CNN

architecture already shows a great ability to generalize to

three independent datasets of videos, while maintaining a

high accuracy that surpasses previous works on this prob-

lem [20]. Moreover, and unlike image processing, the

high accuracy in video classiﬁcation shows no over-ﬁtting,

since all images (one per video) were independent of each

other. Furthermore, its preprocessing allows for a more

rapid, and computationally cheap, classiﬁcation.

6 Future lines

Both image and video classiﬁers can be tried in clini-

cal practice for further enhancement. Image classiﬁcation

showed a lower generalizing ability at ﬁrst. However, the

POCUS database offered a large set of images and the pre-

processing carried out were ﬁnally able to increase its gen-

eralizing ability and showed an acceptable accuracy with-

out compromising the model to overﬁtting. Since the last

results are better, we strongly suggest making a more com-

plex CNN, as well as increasing such database with many

other videos, so that there are much more different frames

and the image classiﬁer can be further enhanced based on

our preprocessing methods and CNN architecture.

On the other hand, the video classiﬁer achieved a high ac-

curacy and was able to generalize, since there were many

different videos from independent sources. All video pro-

cessing was done by student Pablo’s using a very modest

Personal Computer. We believe, nonetheless, that more

complex CNN architectures should be tried out in more

powerful devices -and an even higher accuracy would be

attained. Moreover, further pre-processing could be tested,

although we do recommend maintaining the core idea of

our pre-processing methods, since it was precisely M-

mode US and VGG pre-processing what made it possible,

and boosted video classiﬁcation. In terms of image pro-

cessing we faced the same problem. The resources that we

had were not able to handle a huge amount of images and

complex preprocessing, so the training of the CNN was

limited, as we need to select an speciﬁc batch of each cat-

egory folder that our PC’s could be able to work with.

It is important also to state the simplicity of our approach

to pleural resection. We believe that applying a more ef-

fective way to get the same or even better result would be

optimal, and a way to accomplish that is perhaps building a

segmentation CNN able to determine more accurately the

pleural position and get rid of the actual background points

of the image without even loosing important information.

As a further scope we could face this problem by simply

using powerful resources as Amazon Web Services(AWS)

or Google Cloud Platform(GCP), just setting a sufﬁcient

number of virtual machines with standard CPU power to

help us process the whole amount of POCUS images.

Workload and Acknowledgments

Special thanks to Dr.

Angel (HUFA), LAIMBIO staff

(URJC), Dr. Rafael, and Prof. Cristina, Ph.D., for their

technical, and sometimes personal, support throughout the

whole project.

The paper was written jointly. Each student wrote the fol-

lowing sections:

• Pablo sections Abstract, 1, 2, 3.1, 3.3, 3.5, 4.1.1, 4.2,

5, 6 and References.

• Alejandro sections: 3.4,4.1, 4.1.2, 5, 6.

As for the practical part, there were several collaborators:

• HUFA: Dr.

Angel instructed us on LUS images.

• Dr. Rafael was very helpful and supportive in the ﬁrst

stages of the development of this project.

• LAIMBIO: Both Norberto and David Viar helped us

thanks to their previous work (TFG, or Bachelor’s de-

gree thesis on pneumonia detection, also in HUFA.)

• Student Pablo: First CNN (overﬁtting). POCUS

dataset management: organizing images into folders

(labels), creating frames from the videos, standard

preprocessing, creating splits for cross-validation.

Standard preprocessing for images (normalization,

histogram equalization and mean ﬁlter). Transfer

Learning (unexpected, yet promising results for pneu-

monia). Video preprocessing (standard preprocess-

ing, resizing, VGG preprocessing, M-mode US).

Video classiﬁcation (simple CNN, high accuracy and

generalizing ability).

• Student Alejandro: Complex image preprocessing

(pleural line detection). Complex CNN (overﬁtting).

References

[1] J. Born, N. Wiedemann, M. Cossio, C. Buhre,

G. Br

andle, K. Leidermann, A. Aujayeb, M. Moor,

B. Rieck, and K. Borgwardt, “Accelerating detection

of lung pathologies with explainable ultrasound im-

age analysis,” Applied Sciences, vol. 11, p. 672, Jan

2021.

[2] “Pagano a, numis fg, visone g, pirozzi c, masarone

m, olibet m, nasti r, schiraldi f, paladino f. lung ul-

trasound for diagnosis of pneumonia in emergency

department. intern emerg med. 2015 oct;10(7):851-4.

doi: 10.1007/s11739-015-1297-2. epub 2015 sep 7.

pmid: 26345533.,”

[3] “Buonsenso d, pata d, chiaretti a. covid-19 outbreak:

less stethoscope, more ultrasound. lancet respir

med. 2020 may;8(5):e27. doi: 10.1016/s2213-

2600(20)30120-x. epub 2020 mar 20. pmid:

32203708; pmcid: Pmc7104316..”

[4] “Fiala mj. ultrasound in covid-19: a timeline of ul-

trasound ﬁndings in relation to ct. clin radiol. 2020

jul;75(7):553-554. doi: 10.1016/j.crad.2020.04.003.

epub 2020 apr 18. pmid: 32331781; pmcid:

Pmc7165267..,”

[5] “Amatya y, rupp j, russell fm, saunders j, bales b,

house dr. diagnostic use of lung ultrasound compared

to chest radiograph for suspected pneumonia in a

resource-limited setting. int j emerg med. 2018 mar

12;11(1):8. doi: 10.1186/s12245-018-0170-2. pmid:

29527652; pmcid: Pmc5845910.,”

[6] “Lichtenstein d, goldstein i, mourgeon e, cluzel

p, grenier p, rouby jj. comparative diagnostic per-

formances of auscultation, chest radiography, and

lung ultrasonography in acute respiratory distress

syndrome. anesthesiology. 2004 jan;100(1):9-15.

doi: 10.1097/00000542-200401000-00006. pmid:

14695718.,”

[7] “Bourcier je, braga s, garnier d. lung ultrasound will

soon replace chest radiography in the diagnosis of

acute community-acquired pneumonia. curr infect dis

rep. 2016 dec;18(12):43. doi: 10.1007/s11908-016-

0550-9. pmid: 27785748.,”

[8] “Ant

unez-montes oy, buonsenso d, paz-ortega so.

rationale for the routine application of lung ultra-

sound in the management of coronavirus disease

2019 (covid-19) patients in middle- to low-income

countries. ultrasound med biol. 2020 sep;46(9):2572-

2574. doi: 10.1016/j.ultrasmedbio.2020.05.020.

epub 2020 jun 5. pmid: 32593499; pmcid:

Pmc7274635.,”

[9] “Ragnoli b, malerba m. focus on the potential role of

lung ultrasound in covid-19 pandemic: What more

to do? int j environ res public health. 2020 nov

13;17(22):8398. doi: 10.3390/ijerph17228398. pmid:

33202769; pmcid: Pmc7698284.,”

[10] “Volpicelli g, lamorte a, vill

en t. what’s new in

lung ultrasound during the covid-19 pandemic. in-

tensive care med. 2020 jul;46(7):1445-1448. doi:

10.1007/s00134-020-06048-9. epub 2020 may 4.

pmid: 32367169; pmcid: Pmc7196717.,”

[11] “Kulkarni s, down b, jha s. point-of-care lung ul-

trasound in intensive care during the covid-19 pan-

demic. clin radiol. 2020 sep;75(9):710.e1-710.e4.

doi: 10.1016/j.crad.2020.05.001. epub 2020 may 13.

pmid: 32405081; pmcid: Pmc7218373.,”

[12] S. Ying, S. Zheng, L. Li, X. Zhang, X. Zhang,

Z. Huang, J. Chen, H. Zhao, R. Wang, Y. Chong,

J. Shen, Y. Zha, and Y. Yang, “Deep learning en-

ables accurate diagnosis of novel coronavirus (covid-

19) with ct images,” 2020.

[13] “Hui, d. s. et al. the continuing covid-19 epidemic

threat of novel coronaviruses to global health-the lat-

est 2019 novel coronavirus outbreak in wuhan, china.

int. j. infect. dis. 91, 264–266 (2020).,”

[14] M. Z. Alom, M. M. S. Rahman, M. S. Nasrin, T. M.

Taha, and V. K. Asari, “Covid mtnet: Covid-19 de-

tection with multi-task deep learning approaches,”

2020.

[15] “Smith mj, hayward sa, innes sm, miller asc. point-

of-care lung ultrasound in patients with covid-19 - a

narrative review. anaesthesia. 2020 aug;75(8):1096-

1104. doi: 10.1111/anae.15082. epub 2020 apr 28.

pmid: 32275766; pmcid: Pmc7262296.,”

[16] F. Chollet, “keras.” https://github.com/

fchollet/keras, 2015.

[17] M. Farooq and A. Hafeez, “Covid-resnet: A deep

learning framework for screening of covid19 from ra-

diographs,” 2020.

[18] M. Loey, F. Smarandache, and N. E. Khalifa, “Within

the lack of chest covid-19 x-ray dataset: A novel de-

tection model based on gan and deep transfer learn-

ing,” Symmetry, vol. 12, p. 651, 04 2020.

[19] S. Kulhare, X. Zheng, C. Mehanian, C. Gregory, M.-

H. Zhu, K. Gregory, H. Xie, J. Jones, and B. Wilson,

Ultrasound-Based Detection of Lung Abnormalities

Using Single Shot Detection Convolutional Neural

Networks: International Workshops, POCUS 2018,

BIVPCS 2018, CuRIOUS 2018, and CPM 2018, Held

in Conjunction with MICCAI 2018, Granada, Spain,

September 16–20, 2018, Proceedings, pp. 65–73. 09

2018.

[20] “David viar hern

andez and norberto malpica

gonz

alez. .detecci

on autom

atica de neumon

ıa en

pediatr

ıa usando ultrasonidos.”