Deep Learning and Medical Image Analysis: Epistemology and Ethical

Issues

Francesca Lizzi

1 a

, Alessandra Retico

1 b

and Maria Evelina Fantacci

2 c

National Institute for Nuclear Physics, Pisa Division, Pisa, Italy

Department of Physics, University of Pisa, Pisa, Italy

Keywords:

Deep Learning, Ethics, Epistemology, Multi Disciplinary Science, Medical Imaging.

Abstract:

Machine and deep learning methods applied to medicine seem to be a promising way to improve the perfor-

mance in solving many issues from the diagnosis of a disease to the prediction of personalized therapies by

analyzing many and diverse types of data. However, developing an algorithm with the aim of applying it in

clinical practice is a complex task which should take into account the context in which the software is devel-

oped and should be used. In the ﬁrst report of the World Health Organization (WHO) about the ethics and

governance of Artiﬁcial Intelligence (AI) for health published in 2021, it has been stated that AI may improve

healthcare and medicine all over the world only if ethics and human rights are a main part of its development.

Involving ethics in technology development means to take into account several issues that should be discussed

also inside the scientiﬁc community: the epistemological changes, population stratiﬁcation issues, the opacity

of deep learning algorithms, data complexity and accessibility, health processes and so on. In this work, some

of the mentioned issues will be discussed in order to open a discussion on whether and how it is possible to

address them.

1 INTRODUCTION

Machine and deep learning methods applied to medi-

cal images are proving to be a promising way to im-

prove the performance in solving many issues: the

diagnosis of a speciﬁc disease, the contouring of or-

gans or lesions and the prediction of the prognosis.

Deep learning, in particular, offers the possibility of

analyzing many patients’ data in a reproducible way

and they can be applied to carry out follow up and

radiomic studies. In particular, the advent of Deep

Learning (DL) algorithms in the ﬁeld of medical im-

age analyses is leading to a change in supporting

physicians in their role. Many different applications

have been explored (Litjens et al., 2017) successfully.

However, developing a DL algorithm with the aim of

applying it in clinical practice is a complex task which

should take into account the context wherein the soft-

ware is developed and should be used. In 2021, the

World Health Organization (WHO) published the ﬁrst

report about the ethics and governance of Artiﬁcial

https://orcid.org/0000-0003-0900-0421

https://orcid.org/0000-0001-5135-4472

https://orcid.org/0000-0003-2130-4372

Intelligence (AI) for health (WHO, 2021). In that re-

port, it has been stated that AI may improve health-

care and medicine all over the world only if ethics

and human rights are a main part of its development.

WHO recognizes that ethical guidance based on the

shared perspectives of the different entities that de-

velop, use or oversee such technologies is critical to

build trust in these technologies, to guard against neg-

ative or erosive effects and to avoid the proliferation

of contradictory guidelines. Involving ethics in tech-

nology development means to take into account sev-

eral issues.

In this work some of the most interesting ones

will be discussed. First, understanding how scien-

tiﬁc method is changing should be at least taken into

account and discussed, when developing a medical

software. In fact, most of the ethical issues related

to the application of DL algorithms to clinical prac-

tice are directly connected to the shift of the scien-

tiﬁc paradigm brought by the intense use of data and

data mining. Second, the way we collect data and

build data sets is crucial to develop fair AI-based al-

gorithms. This means to correctly perform the popu-

lation sampling in order to prevent social biases and

to preserve technical information to avoid technolog-

172

Lizzi, F., Retico, A. and Fantacci, M.

Deep Learning and Medical Image Analysis: Epistemology and Ethical Issues.

DOI: 10.5220/0011983000003497

In Proceedings of the 3rd International Conference on Image Processing and Vision Engineering (IMPROVE 2023), pages 172-179

ISBN: 978-989-758-642-2; ISSN: 2795-4943

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

ical biases. In fact, the development of AI-based al-

gorithms should pay attention to the process of im-

age production which includes manufacturers, acqui-

sition parameters and also the interactions with physi-

cians. Moreover, since we should always compare

the predicted results with a ground truth, a labeled

data set collection should consider the inter- and intra-

observer variability as well as the risk of containing

a not negligible label noise (see ﬁg. 1). In this con-

Figure 1: An example of label noise: on the left, an original

lung CT scan (case 0053) from the COVID-19 Challenge

data set (An et al., 2020) is shown with a windowing be-

tween -1000 and 300 HU. On the right: the reference label

mask for a COVID-19 lung lesion published within this data

set is shown in red. The label appears in the axial view (top

row) as a perfect circle probably due to the use of a support

software for image labelling by the radiologist.

text, the application of AI to medical images needs a

special care since its wrong use may harm not only

people but also the healthcare systems (Stevens et al.,

2018).

In this work, the ﬁrst section is dedicated to the

changing scientiﬁc paradigm and on how the so-

called Fourth Paradigm affects studies that concern

deep learning and medical images. In the second sec-

tion, the many issues connected to the collection of a

data set are discussed, including the problem of the

ground truth reliability, of the access to data and their

quality, and the problem of improving the clinical out-

comes in the context of a real hospital protocol.

2 THE IMPORTANCE OF

EPISTEMOLOGICAL CLAIMS

IN HEALTHCARE

The concept of ”scientiﬁc paradigm” was introduced

by Kuhn in the ”Structure of Scientiﬁc Revolutions”

published in 1962 (Kuhn, 1962) and, despite its lim-

its, it is very useful to frame a simpliﬁed scheme of

the evolution of the scientiﬁc paradigms reported by

(Hey et al., 2021), as shown in (Table 1). According

to Kitchin (Kitchin, 2014), we can summarize the idea

of the paradigm according to Kuhn as an accepted

way of interrogating the world and produce knowl-

edge which is common to a substantial proportion of

researcher in a discipline at any one moment in time.

According to Hey (Hey et al., 2021), epistemology

is moving towards a new paradigm called the ”fourth

paradigm” or ”exploration science”. In this evolu-

tion, some fundamental rules of traditional science

are deeply changing and they should be taken into ac-

count since they are useful to establish the limits and

the possibilities of these new rising methods. The as-

sumptions that are made during the development of an

algorithm are critical to deﬁne the model itself and the

boundaries in which it should be applied. According

to Gray (Hey et al., 2021), there are two main ways

to frame the fourth paradigm: the ﬁrst way, typical of

industry, is a pure empiricism wherein machine learn-

ing techniques can reveal the inherent truth of data,

while the second one looks at the fourth paradigm as a

new extension of the established scientiﬁc paradigm.

The fourth paradigm, as pure inductive empiricism

brought by the use of Big Data and also by the ris-

ing of the deep learning methodologies, has the po-

tential to undermine the scientiﬁc legitimacy of the

machine learning (Enni and Herrie, 2021). As an ex-

ample, Campolo and Crawford (Campolo and Craw-

ford, 2020) use the Enchantment theory of Max We-

ber to describe a broader epistemological diagnosis

of modernity. They afﬁrm that not understanding the

motivation that leads a deep learning based model to a

decision could produce the effect of considering that

algorithm as something magical. These considera-

tions do not come only from humanities but also from

“hard” sciences. Stuart J. Russell, a well-known pro-

fessor of computer science from Berkeley University,

in 2018, spoke about deep learning and described it as

“a kind of magic” since we cannot understand when

and why the deep learning hypothesis is consistent.

For this reason, it is interesting to discuss the pro-

cess of knowledge generation, evidence and causa-

tion in particular in the healthcare domain. In Stevens

at al. (Stevens et al., 2018), a critically and health-

care centered review of epistemological claims is pre-

sented. The healthcare ﬁeld is characterized by an

institutionalized set of epistemological principles and

generally accepted scientiﬁc methodologies (Stevens

et al., 2018) (Beltran, 2005) which are challenged by

the deep learning practices. The language used to de-

scribe the applications of algorithm in healthcare can

be an interesting way of analyzing whether there are

Deep Learning and Medical Image Analysis: Epistemology and Ethical Issues

173

Table 1: The evolution of scientiﬁc paradigm according to Hey et al. (Hey et al., 2021).

Paradigm Method Dating

Experimental Science Observation of natural phenomena Pre-Renaissance

Theoretical Science Using models, generalization Pre-Computers

Computational Science Simulations Pre-Big Data

Exploratory Science Data mining Now

different ways of using data in this speciﬁc domain.

Stevens et al., studying systematically the editorials

on the use of Big Data practices in healthcare, de-

scribes ﬁve ideal typical discourses, naming them us-

ing the relations between implicit assumptions about

evidence and knowledge and the diverse epistemolog-

ical positions. The ﬁve categories they design are: the

modernist, the instrumentalist, the pragmatist, the sci-

entist and the critical-interpretative. Despite the de-

tails of each discourse, there is a signiﬁcant differ-

ence between the ﬁrsts four and the last one about

the conceptualization of data: the former consider

data inherently truthful and meaningful (natural and

pre-existing), while the latter consider data as con-

structed and hence they necessarily emphasize some

aspects and leave out others. While a simple positivis-

tic, hypothesis-and-theory free and purely empiricist

approach seems to be a way of making the use of data

simpler in the clinical practice, it is a trap. With re-

gard to the ﬁeld of medical image analysis and con-

sequently the radiological medical domain, we know

that data are not given, natural or pre-existing. Medi-

cal images are the results of:

1. A traditional scientiﬁc process: their production

is based on physical studies on the interaction be-

tween matter and radiation, human body and ra-

diation, on the physical processes of X-ray, mag-

netic ﬁelds and ultrasounds production as regards

radiology, and radioactivity and all the issues

linked to it as regard nuclear imaging;

2. A technological development history: the im-

age production deals with the detectors improve-

ments, the materials used for detecting photons,

the electronics which, simplifying, determine for

example the spatial resolution, the contrast and

other image quality characteristics. When we

deal with 3D images such as Computed Tomog-

raphy (CT) scans we should consider the image

reconstruction algorithms which are a mix of tra-

ditional scientiﬁc research, especially mathemati-

cal one (for example Radon transform or Fourier

signal analysis) and pure technological improve-

ments such as the sliding contacts.

3. An industrial process: medical imaging systems

are not equally distributed around the world and

their production is highly costly producing as a

result that there are few vendors that deals with the

imaging machinery market. Moreover, as a result

of an industrial process, some parts of the medical

images production are protected by patents which

inhibits the complete knowledge on how an image

is produced;

4. A function-based process: medical images are

made on the basis of their utility and improved

following their possible uses in hospitals. The

choice of using a speciﬁc imaging modality de-

pends on the scope (morphology or functionality

and diagnosis, follow up, radiotherapy planning,

...) and on the part of the body that needs to be

imaged. They are made to be presented to physi-

cians in a way that medical doctors can interpret

and taking into account the speciﬁc medical for-

mation process they attended. Moreover, contrar-

ily to natural images, most of medical imaging

modality implies the delivering of a radiation dose

to the patients, making their use a dynamic equi-

librium between costs and risks, in terms of capi-

tal and health, and beneﬁts.

For all these reasons, it is unacceptable to consider

medical images data as pre-existing or natural. More-

over, applying Big Data techniques, such as machine

and deep learning, cannot be considered as a free hy-

pothesis science. Even if the hypothesis is a complex

hypothesis and it is very far from the pre-Renaissance

way of formulating it, we are always assuming that,

given the constructed data and the context, there is

a model which may solve the given task we want to

study. This means that we are using data that con-

tain already the solution. Characterizing the machine

and deep learning techniques as comprehensive and

intrinsically unbiased can be misleading rather than

helpful in shaping scientiﬁc as well as public percep-

tions of the features, opportunities and dangers asso-

ciated with data-intensive research (Leonelli, 2014).

Finally medical images can rarely be considered as

Big Data and hence the application of techniques de-

veloped on them should be even more careful as re-

gards, for example, the generalization goal. What is

at stake is our ability to produce knowledge not only

in a traditional scientiﬁc way but making it from a

critical position, avoiding the accusation of practising

”magic” or ”alchemy”. It deals with knowing and as-

suming how much complex is to create algorithms,

IMPROVE 2023 - 3rd International Conference on Image Processing and Vision Engineering

174

especially deep learning one, with the scope of ap-

plying them in hospitals and, by assuming it, proceed

towards a fair, scientiﬁc, active and impacting appli-

cation of machine and deep learning to medicine.

3 BUILDING MEDICAL IMAGE

DATA SETS

As discussed in the previous section, medical data sets

of images cannot be considered inherently truthful as

natural resources. It is interesting to discuss the issues

that should be considered in a data set collection that

demonstrates the thesis of the ﬁrst section. In particu-

lar, the ground truth production and accessibility and

quality of medical images data sets will be considered

in the following.

3.1 Ground Truth Reliability

The ground truth on medical images is usually made

by medical doctors opinions or by a consensus among

them. For this reason, it always suffers from a certain

grade of variability which should be kept under con-

trol (Bridge et al., 2016) (Renard et al., 2020). The

use of a peculiar imaging modality and of a speciﬁc

imaging system may affect the capability of having a

reliable ground truth and the aggregation of data com-

ing from different sources is a challenge that still need

to be addressed. In fact, publicly available data sets of

medical images usually are small data sets that need to

be aggregated in order to obtain a set with a sufﬁcient

number of samples to train a deep learning algorithm

(Lizzi et al., 2021). Even if it is possible to collect

private image data sets from hospitals, the process is

very time consuming for both the collection and the

labeling. Moreover, the publication of such data sets

may not be possible, thus reducing the chance of re-

producing results obtained by other studies. Publish-

ing the data is not easy because database maintenance

is expensive and the privacy of patients has to be man-

aged rightfully. The ground truth usually depends on

the task we want to solve and its creation is a pivotal

step for algorithm development. If the task we want to

solve is a classiﬁcation task, the ground truth consists

in assigning a class to each image or patient in the

data set. A patient image taken at different time points

may belong to a different class because human body

changes over time. The way the classes are deﬁned is

mainly based on medical protocols but medical pro-

tocols do not guarantee the variability delete. Even if

this labelling process seems to be very fast, when a

huge amount of labelled data is required, the process

is very time consuming for doctors. Another way for

labelling medical images is to assign to each pixel or

each voxel a certain class. This kind of labelling is

suited for solving segmentation problems. A medi-

cal image usually contains many pixel and voxel and

this characteristic makes the labelling extremely time

consuming. If we suppose to have a standard lung

CT scan with size 512x512x100 the number of ele-

ments to be labelled is more than 26 millions! There

are some tools to help physicians in this task but they

may introduce a bias in the labelling. In order to re-

duce the cost of labelling, the use of non-expert peo-

ple has been employed in the ﬁeld of natural images

(Kuznetsova et al., 2020), but the use of such kind

of labelling process leads to highly noise data sets.

In the medical images domain, in which the objects

to be identiﬁed are usually small and are difﬁcult to

be identiﬁed, this process is even harder. Having the

availability of large labeled data sets of medical im-

ages is currently a real challenge even despite the la-

belling process. Medical images data sets, in fact, are

usually small and their collection is not easy because

of privacy issues and institutional policies.

3.2 Data Accessibility and Quality

Data may come from public or private collections.

Both of these two modalities have weaknesses and

strengths which are going to be discussed in the fol-

lowing. First of all, public data may be effectively

public, i.e. accessible to every one, or they may be

accessed through a speciﬁc agreement. Private data

are instead data which can not be used or accessed

in any case. Data may be not accessible for many

reasons. One of the most problematic is related to

privacy. In order to better understand the risks of pub-

lishing data, it is interesting to discuss the most used

image formats. This is important because medical im-

age formats usually contain a header with patient and

physician information. There are mainly two data

formats typically used for medical images and they

are the Neuroimaging Informatics Technology Initia-

tive (NIfTI) and the Digital Imaging and Commu-

nications in Medicine (DICOM) (Standard DICOM,

2021). The NIfTI format was created in the ﬁeld of

neuroimaging and it is a standard which contains a

header with only information about orientation, voxel

size and image visualization. 3D images, for example

CT scans or MRI scans, can be stored in this format

which deﬁnes uniquely the correct orientation and the

physical volumes. In Figure 2, an example of a NIfTI

header of a CT scan is reported.

The Digital Imaging and Communications in

Medicine (DICOM) standard (Standard DICOM,

2021) is the global convention used by manufacturers

Deep Learning and Medical Image Analysis: Epistemology and Ethical Issues

175

Figure 2: An example of a whole NIfTI header. It can be

noticed that this header includes only information about the

voxel size and the image orientation.

to deﬁne and store diagnostic imaging data. DICOM

images are encoded as a set of elements; public ele-

ments are deﬁned by the DICOM standard, and pri-

vate elements are deﬁned on an individual basis by

each manufacturer. A DICOM data element or at-

tribute is made of 3:

• a tag that identiﬁes the attribute, usually in the for-

mat (XXXX, YYYY) with hexadecimal numbers,

and may be divided further into DICOM Group

Number and DICOM Element Number;

• a DICOM Value Representation (VR) that de-

scribes the data type and format of the attribute

value.

The ﬁelds of the DICOM header contain many

information from the patient ID, which is a number

that uniquely identiﬁes the patient, the Patient’s Birth

Name (0010,1005), the Patient’s Age (0010,1010),

the Patient’s Size (0010,1020), the Patient’s Address

(0010,1040) or even the Patient’s Mother’s Birth

Name (0010,1060). All these data are a problem when

we deal with privacy because they may allow a com-

plete re-identiﬁcation of subjects. On the other hand

the DICOM format contains also information on the

acquisition parameters such as the reconstruction ker-

nel, the imaging system used, exposure time, X-ray

tube current, the ﬁeld of view (FOV) size or the re-

constructed FOV.

These characteristics are less prone to be problem-

atic with regard to privacy and they are very useful for

algorithms development. However, in most of pub-

lished medical images data all this information is lost.

This is mainly due to the fact that it is not so easy to

Figure 3: An example of a part of a DICOM header. Con-

trarily to the NIfTI one, the DICOM header contains infor-

mation about the patient, the instrumentation, the acquisi-

tion protocol and the clinical personnel.

treat privacy and DICOM standard, since the number

of tags that may be contained is very large. Moreover,

making studies on humans implies not only privacy

related issues but ethical issues too. For these rea-

sons accessing to Italian hospital data requires a strict

protocol to be carried out. The image instrumentation

manufacturers use private elements to encode acqui-

sition parameters that are not yet deﬁned by the DI-

COM standard or that they consider proprietary. They

also deﬁne and include private elements that contain

Protected Health Information (PHI). These PHI pri-

vate elements can be as obvious as the name of a pa-

tient and as subtle as an identiﬁer string that could be

tracked back to a patient by someone with access to

the departmental image archive. A DICOM confor-

mance statement is a document published by a manu-

facturer that contains technical information concern-

ing data exchange with a speciﬁc type of device (e.g.

an imaging unit, workstation, printer, image archive).

The conformance statement provides the mechanism

for a manufacturer to publish the set of private ele-

ments that are stored in the DICOM ﬁles created by

an imaging system. Manufacturers do not document

and publish all of their private elements. For these

reasons, the de-identiﬁcation process should meet two

conﬂicting requirements: (i) any PHI must not be in-

cluded in exported data and (ii) the system must retain

all data that describe the acquisition, such as physical

parameters for individual images, as well as other pa-

rameters such as series description. De-identifying a

IMPROVE 2023 - 3rd International Conference on Image Processing and Vision Engineering

176

DICOM image is a challenging task that carries the

risk of leaving in the header PHI or meta-data that

makes the re-identiﬁcation possible. On the other

hand, the NIfTI format has been invented to have not

patient information in the header but it does not allow

to store important technical parameters. It could be

interesting to study a new image format standard suit-

able for AI and deep learning algorithms which con-

tains all the technical information while keeping the

privacy risk as lower as possible. The availability of

acquisition parameters is crucial to have a high quality

data set and the research of a good trade-off between

accessibility and quality is an urgent challenge to be

tackled.

3.3 How to Gain Impact?

Over the last 10 years, publications on AI in radiology

have increased from 100–150 per year to 700–800

per year (Pesapane et al., 2018) and the interest in

the medicine ﬁeld is continuously increasing. AI and

deep learning studies mainly focus their scopes on in-

creasing the accuracy of diagnosis when compared to

the physicians performances.

However high accuracy does not necessarily mean

that an AI algorithm improves clinical outcomes. It

is, in fact, important to assess whether its use in clin-

ical practice can be integrated in the hospital work-

ﬂow and how much the impact is, not only on the

outcomes, but also on the physicians’ training. In or-

der to perform this kind of analysis, a clinical trial

is needed and clinical trial studies are usually time

consuming and expensive. It is pivotal to question

which could be the role of AI in the medical and clin-

ical workﬂow, especially in the radiology ﬁeld which

seems to be the most explored ﬁeld of medicine. It

is also interesting to discuss the role of a radiologist

in the hospital workﬂow and whether they can be re-

placed by an artiﬁcial intelligence or be supported

by it. In Pesapane et al. (Pesapane et al., 2018), a

group of radiologists reﬂects on what it means to let

an AI make a diagnosis and what are the differences

between the human evaluation and the AI one.

AI and especially deep learning functioning in ra-

diology is based on a principle that is very similar to

the clinical one: “the more images you see, the more

examinations you report, the better you get” and this

may be the reason why AI is successfully applied to

radiology. Since the comparison between the radiol-

ogist’s and AI performance depends on the radiolo-

gist experience and also on the quality of the devel-

oped AI, it is not straightforward to state whether and

when one is better than the other. When image anal-

ysis takes too much time with respect to the neces-

sity of the patient, i.e. a very urgent clinical evalua-

tion is necessary, AI may be very helpful in a hospital

workﬂow. As an example, in this study (Kim et al.,

2021), the application of a deep learning-based assis-

tive technology in the Emergency Department (ED)

context has been studied on Chest Radiographs (CRs).

CR interpretation is a difﬁcult task that requires both

experience and expertise because various anatomical

structures tend to overlap when captured on a single

two-dimensional image, different diseases may have

a similar presentation and speciﬁc diseases may be

present with different characteristics. For these rea-

sons, the CR interpretation suffers from a signiﬁcant

possibility of misinterpretation, reaching the 22% ac-

cording to (Donald and Barnard, 2012). ED physi-

cians perform worse than trained radiologists in read-

ing images. However, radiologists may not be avail-

able, especially during nights and weekend and CR

interpretation in the ED settings is given to ED physi-

cians. For all these reasons, Kim et al. (Kim et al.,

2021) studied whether an ED physician supported by

a deep-learning based algorithm for CR interpretation

performs better than the single ED physician. They

found that ED departments may beneﬁt from the use

of AI even if this experiment needs at least an exter-

nal validation study. This is an example that shows

clearly how much it is important to know the health-

care domain and practice in order to structure a deep

learning experiment.

Despite the improvements deep learning may pro-

duce to healthcare, another pivotal question concerns

the problem of accountability. When an AI is used to

make a decision in clinical practice, it is not trivial to

understand who is responsible for the diagnosis. In

this work (Neri et al., 2020), a radiologist supported

by an AI is depicted as responsible for the diagnosis

if they are trained on the use of AI since they are re-

sponsible for the actions of machines. Moreover, it

is necessary to deepen the research ﬁeld of explain-

ability in order to let the radiologists understand the

AI behaviour. Furthermore, the use of AI may bias

the radiologist decisions. Lastly, even the public dis-

cussion on the introduction of AI systems as possible

substitutes of the physicians themselves can be dan-

gerous and produce a paradox effect: since radiolo-

gists are going to be replaced by AI, there will be a

lack of motivation for young doctors to pursue a ca-

reer in radiology.

For all these reasons, building and even bringing

in the public debate deep learning models to be ap-

plied to radiology is a delicate task and claiming clin-

ical advancements without clinical trials is unfair. As

described above, medicine and healthcare ﬁelds are

complex domains and the advancement that comes

Deep Learning and Medical Image Analysis: Epistemology and Ethical Issues

177

from the introduction of deep-learning based tech-

nologies should be carefully evaluate inside that con-

text to evaluate the real impact. Including information

on context at the beginning of the development of a

DL algorithm could help to gain clinical impact.

4 CONCLUSIONS AND

DISCUSSION

Some of the many aspects that deal with the creation

of a deep learning algorithm applied to medicine and,

speciﬁcally, to imaging have been discussed. The dif-

ﬁculty of taking into account all the issues is clear and

they relate to many ﬁelds of knowledge.

In section 2, the changing scientiﬁc paradigm

has been discussed. How researchers pose their re-

search questions and which epistemological assump-

tions they embrace are fundamental to understand the

kind of research they are doing. This process cannot

be done without looking at the social processes that

leads to the data collection and the data generation.

In section 3.2, the process of collecting a data set

and subsequently building a ground truth on medi-

cal images is discussed with potentialities and limi-

tations. Typically, the ground truth on medical im-

ages is made by the physician opinion or by a consen-

sus among many medical doctors. When made with

the second modality, the ground truth always suffers

from a variability that is difﬁcult to be erased. The

quality of an algorithm strictly depends on the qual-

ity of the ground truth but having a large number of

physicians is economically expensive and requires a

high grade of coordination and collaboration among

research and health institutes. The quality of the al-

gorithms depends also on the quality of data that can

be private or public. Public data guarantees the possi-

bility of testing different algorithms on the same data

set but, in order to make them publishable, important

information on, for example, acquisition protocols or

scanners, may be lost. Private data has the advantage

to be designed for the speciﬁc experiment and taken

following inclusion criteria decided by the collector.

When released, this kind of data can be designed to

contain the information on acquisition that could be

useful and meaningful for the analysis. In any case,

medical images data are scarce and they may lack of

label quality. This issue is one of the most limiting in

deep learning algorithm development. Lastly, when

an algorithm is developed to be used in clinical prac-

tice, it has to be validated not only statistically but

also clinically. The performances on a test set are

not sufﬁcient to claim for clinical advancements since

it belongs to the same data set used for the training

and the validation. The algorithm, in fact, needs to

be tested also on at least an independent external data

set to evaluate its generalization capability. The ex-

ternal data should be taken from another medical cen-

ter and should contain the information on acquisition

and scanners in order to make possible the analysis

of the image characteristics that may confuse the al-

gorithm. This process can be done in two modali-

ties, case-control and clinical trial studies, and both

of them suffer from the issues to correctly represent

the population.

Once the algorithm has been externally validated,

it should be integrated in the hospital workﬂow and

its performance should be evaluated also in this con-

text. It has been established that the capacity of an

algorithm to outperform a physician is strictly con-

nected to the experience of the physician to solve that

speciﬁc task. For this reason, there exist situations in

which the application of an algorithm may be really

helpful to both increase performance and save time.

In this context, it is interesting to question who is re-

sponsible for the diagnosis when an algorithm is used

to support physicians or directly to diagnose a certain

disease. In order to solve this issue, we need juridical

instruments that helps the application of algorithms in

clinical practice. Building responsibility means also

to train physicians to the use of AI in order to make

them mindful of its use and to produce an informed

consent that patients can really understand.

All these issues suggest that the development of

an algorithm for clinical applications need a deep

and widespread knowledge in all of the cited ﬁelds:

medicine, radiology, healthcare processes, laws, com-

puter science, computer engineering, physics, social

sciences, philosophy and so on. We believe that the

future research on the mentioned issues should pro-

ceed along two paths: on one hand, we need to think

and implement controlled experiments in order to sys-

tematically understand, for example, how the physical

parameters of images, such as kVp or reconstruction

algorithms, affects the feature learning of DL mod-

els; on the other hand, we should study how to build

high quality shareable inclusive privacy-preserving

data sets, not only as a benchmark for performance

comparison but also for the whole algorithm develop-

ment. This could be done, for example, by creating a

new medical image standard that considers the use of

data for scientiﬁc analysis and purpose as an intrinsic

property of the standard itself.

What is at stake is to develop a high-performance,

inclusive and trustworthy AI.

IMPROVE 2023 - 3rd International Conference on Image Processing and Vision Engineering

178

REFERENCES

An, P., Xu, S., Harmon, S. A., Turkbey, E. B., Sanford,

T. H., Amalou, A., Kassin, M., Varble, N., Blain,

M., Anderson, V., Patella, F., Carraﬁello, G., Turkbey,

B. T., and Wood, B. J. (2020). CT Images in COVID-

19.

Beltran, R. A. (2005). The Gold Standard: The Challenge

of Evidence-Based Medicine and Standardization in

Health Care. Journal of the National Medical Associ-

ation, 97(1):110.

Bridge, P., Fielding, A., Rowntree, P., and Pullar, A.

(2016). Intraobserver Variability: Should We Worry?

Journal of Medical Imaging and Radiation Sciences,

47(3):217–220.

Campolo, A. and Crawford, K. (2020). Enchanted Deter-

minism: Power without Responsibility in Artiﬁcial In-

telligence. Engaging Science, Technology, and Soci-

ety, 6:1–19.

Donald, J. J. and Barnard, S. A. (2012). Common patterns

in 558 diagnostic radiology errors. Journal of Medical

Imaging and Radiation Oncology, 56(2):173–178.

Enni, S. A. and Herrie, M. B. (2021). Turning biases into

hypotheses through method: A logic of scientiﬁc dis-

covery for machine learning. Big Data and Society,

8(1).

Hey, T., Tansley, S., and Tolle, K. M. (2021). Fourth

Paradigm.

Kim, J. H., Han, S. G., Cho, A., Shin, H. J., and Baek,

S.-E. (2021). Effect of deep learning-based assistive

technology use on chest radiograph interpretation by

emergency department physicians: a prospective in-

terventional simulation-based study. BMC Medical

Informatics and Decision Making, 21(1):1–9.

Kitchin, R. (2014). Big Data, new epistemologies and

paradigm shifts. Big Data and Society, 1(1):1–12.

Kuhn, T. S. (1962). The structure of Scientiﬁc Revolution,

volume I,II.

Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin,

I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M.,

Kolesnikov, A., Duerig, T., and Ferrari, V. (2020). The

Open Images Dataset V4: Uniﬁed Image Classiﬁca-

tion, Object Detection, and Visual Relationship De-

tection at Scale. International Journal of Computer

Vision, 128(7):1956–1981.

Leonelli, S. (2014). What difference does quantity make?

On the epistemology of Big Data in biology. Big Data

and Society, 1(1):1–11.

Litjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A.,

Ciompi, F., Ghafoorian, M., van der Laak, J. A., van

Ginneken, B., and S

anchez, C. I. (2017). A survey

on deep learning in medical image analysis. Medical

Image Analysis, 42(December 2012):60–88.

Lizzi, F., Brero, F., Cabini, R. F., Fantacci, M. E., Piffer, S.,

Postuma, I., Rinaldi, L., and Retico, A. (2021). Mak-

ing data big for a deep-learning analysis: Aggrega-

tion of public COVID-19 datasets of lung computed

tomography scans. Proceedings of the 10th Interna-

tional Conference on Data Science, Technology and

Applications, DATA 2021, (Data):316–321.

Neri, E., Coppola, F., Miele, V., Bibbolino, C., and Grassi,

R. (2020). Artiﬁcial intelligence: Who is responsible

for the diagnosis? Radiologia Medica, 125(6):517–

521.

Pesapane, F., Codari, M., and Sardanelli, F. (2018). Artiﬁ-

cial intelligence in medical imaging: threat or oppor-

tunity? Radiologists again at the forefront of innova-

tion in medicine. European Radiology Experimental,

2(1).

Renard, F., Guedria, S., Palma, N. D., and Vuillerme, N.

(2020). Variability and reproducibility in deep learn-

ing for medical image segmentation. Scientiﬁc Re-

ports, 10(1):1–16.

Standard DICOM (2021). DICOM standard.

Stevens, M., Wehrens, R., and de Bont, A. (2018). Con-

ceptualizations of Big Data and their epistemological

claims in healthcare: A discourse analysis. Big Data

and Society, 5(2):1–21.

WHO (2021). Ethics and Governance of Artiﬁcial Intelli-

gence for Health Ethics and Governance of Artiﬁcial

Intelligence for Health 2.

Deep Learning and Medical Image Analysis: Epistemology and Ethical Issues

179