*Commentary* **Artificial Intelligence in Digital Pathology: What Is the Future?** *Part 1: From the Digital Slide Onwards*

**Maria Rosaria Giovagnoli <sup>1</sup> and Daniele Giansanti 2,\***


**Abstract:** This commentary aims to address the field of *Artificial intelligence* (AI) in *Digital Pathology* (DP) both in terms of the global situation and research perspectives. It has four polarities. *First*, it revisits the evolutions of digital pathology with particular care to the two fields of the digital cytology and the digital histology. *Second*, it illustrates the main fields in the employment of AI in DP. *Third*, it looks at the future directions of the research challenges from both a clinical and technological point of view. *Fourth*, it discusses the transversal problems among these challenges and implications and introduces the immediate work to implement.

**Keywords:** e-health; medical devices; m-health; digital-pathology; picture archive and communication system; artificial intelligence; cytology; histology

#### **1. Introduction**

Diagnostic pathology has undergone important changes and leaps forward by means of digitalization, which have allowed, from time to time, on the one hand, important changes in decision-making processes, and on the other, important changes in *workflow* and therefore in the *job description* of the insiders [1,2].

All this has had an important impact on the organization of work from one side and on the training of the figures involved in the activities on the other, having to prepare them to make the necessary changes to adapt them to the ever-changing *job description* and interactions with the tools (optics/mechatronics/informatics) in ever-more rapid obsolescence and gradually being more and more able to integrate with *eHealth* and *mHealth* [1–6].

We are moving from physical storage systems of slides to virtual storage of virtualslides (i.e., *e-slides* or *digital-slides*) [3].

Old problems such as the organization of physical storage spaces are giving way to new problems such as physical (conservative) data security and cybersecurity.

Now there is less talk of archives and multi-archives for slides and more and more of how many petabytes or exabytes will be needed for the *e-slides*.

The changes have been so rapid that someone is starting to ask the fateful question: Will the microscope still be needed as we know it today?

We can undoubtedly highlight how, to date, diagnostic pathology has gone through two important revolutions.

The great *innovations* in the field of the diagnostic pathology involved first the introduction of the immune-histo-chemistry in 1980 and second in the introduction of *next-generation sequencing for cancer diagnostics* around 2010.

The *first revolution* involved the introduction of digital pathology and therefore of the key elements from the *e-slide*, up to the acquisition system (video-camera or scanner) and to archiving system, the picture archive and communication system (PACS) for digital pathology [3].

**Citation:** Giovagnoli, M.R.; Giansanti, D. Artificial Intelligence in Digital Pathology: What Is the Future? *Part 1: From the Digital Slide Onwards*. *Healthcare* **2021**, *9*, 858. https://doi.org/10.3390/ healthcare9070858

Academic Editor: Jitendra Singh

Received: 14 June 2021 Accepted: 4 July 2021 Published: 7 July 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

This *second revolution*, if we leave out the era of robotic telepathology (which does not seem to have had a great impact in pathological diagnostics), had two important moments that we can call (a) the *revolution of digital pathology in eHealth* [5] with the possibility of accessing from the personal computer to PACS servers through virtual microscopy and (b) the *revolution of digital pathology in mHealth* [1] with the possibility of accessing the same servers from smartphones and tablets through a virtual microscope. As it has been highlighted by M Avanzo et al. in the review [6], nowadays, AI shows (1) the potentiality to access and correlate large amount of data and (2) direct prospective in the world of diagnostics.

Regarding (1), it is highlighted that today [6], both radiological and pathology images are stored in the PACS; moreover, with the introduction of electronic health records (EHRs), systematic collections of patient health information have been made available, which include both qualitative data, medical records, and laboratory and diagnostics information. AI, if applied to these big digital stores, could prove useful for epidemiological, clinical, and research studies.

Scientists in DP could benefit of AI from combining histopathological data obtained, analyzed, and shared with other sources of clinical data such as that obtained from omics and/or other databases with clinical data/demographic data and/or sources with BIG-DATA.

With regard to (2), it is highlighted that the development of the digital pathology [6] due to the introduction of whole-slide scanners and the progression of computer vision algorithms have significantly grown the usage of AI to perform tumor diagnosis, subtyping, grading, staging, and prognostic prediction. In the big-data era, the pathological diagnosis of the future could merge proteomics and genomics.

It is evident that AI is clearly helping to integrate information from multiple sources. Furthermore, neural networks from AI are used, for example, to extract pertinent details from written notes from the slide representation.

In general, all of us are also expecting AI in DP as a deus ex machina to diminish the error rate and optimize the time of work.

#### **2. Purpose**

The contribution is in line with the Special Issue "The Artificial Intelligence in Digital Pathology and Digital Radiology: Where Are We?" https://www.mdpi.com/journal/ healthcare/special\_issues/AI\_Digital\_Pathology\_Radiology (6 July 2021) [6].

The aim is to highlight, in light of the foregoing, the important aspects of the transitions towards DP and AI, highlighting: (a) the lights and shadows relating to the introduction of AI based on DP and (b) what could be the future directions to face to stabilize the AI in DP.

#### **3. The Revolution of the Digital Slide**

The introduction of *digital slides* (*e-slide* or *virtual-slide*) is undoubtedly a revolutionary change for the pathologist, comparable to that of the introduction of *google maps* for *cartography*.

Through digital pathology, it is in fact possible to navigate through the *e-slide* with reference to coordinates, perform *Zoom and Pan* operations and set references just as with *Google Maps*. Historically, DP in the first applications was faced with implementing telepathology connections [3]. In the first phases, there was talk of *telepathology* and not of DP. Conceptually, there were and still there are two methods to face *telepathology* (TP): *static TP* and *dynamic TP*. *Static TP* consists of the capture and digitalization of images selected by a pathologist or pathologist assistant, which are then transmitted remotely through electronic means. *Dynamic TP* consists of the direct communication between two different centers by using microscopes equipped with a *tele-robotic system* oriented to explore the slide, remotely operated by the *tele-pathologist* or an *assistant tele-pathologist* to reach a remote diagnosis. As an alternative solution between the two methods, widely increased, year after year, there is the *virtual microscopy* (*VM*) starting from the first applications. The latter does not refer to the *tele-control of microscopes*, whilst the glass is scanned as a whole, producing an *e-slide*, and a pathologist or the assistant pathologist can navigate remotely

(via internet) inside this *e-slide* or *virtual slide* in a manner akin to a real microscope. It has to be considered that a single file representing the *e-slide* for pathology applications could reach several tens of gigabytes, more than in the case of applications of digital echography. Thus, the design of an appropriate visualization strategy is a basic core aspect.

Today, the diffusion of the VM was helped by: (a) the availability of fast internet connections; (b) the availability of consolidated visualization strategies; (c) the availability of power image acquisition cameras/scanners; (d) the availability of free visualization software.

We can clearly consider today that VM is an integral part of DP. Therefore, it can be used in biomedical laboratories with great potential. This can affect the organization of work and has the potential to change and improve training [2,4].

DP is not only *digital slides* [6]. However, it is impossible not to point out that *digital slides/e-slides* are a large part of DP.

For this reason, it is important to highlight some strategic aspects of this discipline of the VM and to consider how they evolved over the time.

#### *3.1. The Difference between the Digital Cytology and Digital Histology*

The cytologist and the histologist interact differently with the slides; therefore, when moving to the digital world, this aspect must be strongly considered. The cytologist analyses the cell while the histologist analyses the tissue. If we can make a comparison with architecture, the cytologist focuses on the brick and looks inside, whilst the histologist looks at the entire wall. For the cytologist to look inside the cell, it is particularly important to use the focus function, which is not needed by the histologist. This translates into cytology in an important need for digitization: that of allowing the focus function in the digital world. This is implemented with the creation of different digital layers to simulate fire through the *Z-stack* [3] function or other solutions that currently do not allow automatic implementation [7]. For these reasons, the *e-slide* in cytology requires an exorbitant memory occupation to cope with the *Z-stack*.

#### *3.2. The Two Steps of the Revolution of the Digital Pathology: Integration into eHealth and mHealth*

When we refer to the introduction of digital pathology, we must duly consider that there have been two important phases synchronized with the evolution of ICT that in healthcare have led to the developments of *eHealth* and *mHealth* applications.

Consequently, the first *client-server* informatic buildings had, in the era of *eHealth* developments, a strong component based on architectures based on PCs that connected via LAN/WAN.

Figure 1 shows an example of PC access to a *virtual slide* in the case of digital cytology. Subsequently, starting with the release in 2008 of the first smartphones and/or tablets

as we know them today [1], digital pathology has begun to find a fertile vehicle in *mHealth*. Figure 2 highlights a first application in *mHealth* in digital cytology with the Nokia c6 with the operative system Symbian (Symbian Ltd., Southwark UK) device, a border element between mobile phones and smartphones in a WI-FI hotspot.

Figure 3, again with reference to digital cytology, reports some accesses in *mHealth* by a tablet (A), from a train without WI-FI, and in other situations via smartphone (B).

*Healthcare* **2021**, *9*, x FOR PEER REVIEW 4 of 13

*Healthcare* **2021**, *9*, x FOR PEER REVIEW 4 of 13

**Figure 1.** Access to the digital slides using eHealth. **Figure 1.** Access to the digital slides using eHealth. **Figure 1.** Access to the digital slides using eHealth.

**Figure 2.** Access to the digital slides using mHealth not using the smartphone. **Figure 2.** Access to the digital slides using mHealth not using the smartphone.

*Healthcare* **2021**, *9*, x FOR PEER REVIEW 5 of 13

**Figure 3.** Access to the digital slides using mHealth (**A**) while navigating by a train without Wi-Fi; (**B**) a static connection. **Figure 3.** Access to the digital slides using mHealth (**A**) while navigating by a train without Wi-Fi; (**B**) a static connection.

#### *3.3. The Acceptance of the Introduction: The HTA Studies Based on Properly Designed Surveys 3.3. The Acceptance of the Introduction: The HTA Studies Based on Properly Designed Surveys*

A strategic aspect in the introduction of a technology is that of acceptance. Important aspects can be overlooked; moreover, problems of interaction with technologies that depend on generations could also arise. For example, the cytologist while navigating with the traditional microscope has a way of navigating and noticing important details with the side of the eyeball facing outward like that of primitive man to protect himself from attacks by ferocious predators. Switching to a PC-based method in *eHealth* first and *mHealth* on a smartphone or tablet later determines a radical change. Therefore, it is necessary to carry out targeted studies on the acceptance of technologies, focused on the actors, with reference to the most critical applications, as in the case of digital cytology. In the study reported in [5], we highlighted the importance of a health technology assessment approach based on a survey centred on the figures involved from a working point of view in digital cytology (which, as we have seen, presents major problems) in the *eHealth* phase. In the study reported in [1], we highlighted the importance of a health technology assessment approach with a similar configuration in the *mHealth* phase. A strategic aspect in the introduction of a technology is that of acceptance. Important aspects can be overlooked; moreover, problems of interaction with technologies that depend on generations could also arise. For example, the cytologist while navigating with the traditional microscope has a way of navigating and noticing important details with the side of the eyeball facing outward like that of primitive man to protect himself from attacks by ferocious predators. Switching to a PC-based method in *eHealth* first and *mHealth* on a smartphone or tablet later determines a radical change. Therefore, it is necessary to carry out targeted studies on the acceptance of technologies, focused on the actors, with reference to the most critical applications, as in the case of digital cytology. In the study reported in [5], we highlighted the importance of a health technology assessment approach based on a survey centred on the figures involved from a working point of view in digital cytology (which, as we have seen, presents major problems) in the *eHealth* phase. In the study reported in [1], we highlighted the importance of a health technology assessment approach with a similar configuration in the *mHealth* phase.

#### *3.4. The Potentialities in the e-Learning/Remote Training*

*3.4. The Potentialities in the e-Learning/Remote Training*  There is no one who does not see, in the COVID-19 era, that DP has important advantages in training regarding social distancing and the lightening of laboratories. Today, it is possible to access large databases and select targeted e-slide-based studies. Just to give an example, Leeds also has important archives with free access to the site https://www.vir-There is no one who does not see, in the COVID-19 era, that DP has important advantages in training regarding social distancing and the lightening of laboratories. Today, it is possible to access large databases and select targeted e-slide-based studies. Just to give an example, Leeds also has important archives with free access to the site https://www.virtualpathology.leeds.ac.uk/, accessed on 6 July 2021, [8].

tualpathology.leeds.ac.uk/ [8]. See one of the many studies directly navigable with your browser in *eHealth* or *mHealth* by accessing the dedicated archive https://www.virtualpathology.leeds.ac.uk/slides/library/ [9], having fun with one of the many digital slides when navigating using a virtual microscope and simple mouse clicks https://www.virtualpathology.leeds.ac.uk/slides/library/view.php?path=%2FResearch\_4%2FTeach-See one of the many studies directly navigable with your browser in *eHealth* or *mHealth* by accessing the dedicated archive https://www.virtualpathology.leeds.ac.uk/ slides/library/, accessed on 6 July 2021, [9], having fun with one of the many digital slides when navigating using a virtual microscope and simple mouse clicks https://www. virtualpathology.leeds.ac.uk/slides/library/view.php?path=%2FResearch\_4%2FTeaching% 2FEducation%2FManchester\_FRCPath%2FDN%2F124388.svs, accessed on 6 July 2021, [10].

ing%2FEducation%2FManchester\_FRCPath%2FDN%2F124388.svs [10]. In teaching, we highlighted the possibility of setting two important approaches [2]: (a) that of using very large tablets such as LIMS whiteboards or other ones in a finger-In teaching, we highlighted the possibility of setting two important approaches [2]: (a) that of using very large tablets such as LIMS whiteboards or other ones in a fingerbased and cooperative way to navigate virtual slides (Figure 4A), and (b) the other one

based and cooperative way to navigate virtual slides (Figure 4A), and (b) the other one

sharing.

[6].

standard in DP.

based on a slide viewer or scope system with a webcam and a network transmitter to tablet/smartphone, even when not present (Figure 4B), such as, for example, the DMshare system (Leica Microsystems Co., Nussloch GmbH, Germany). Both have allowed to free up important resources in this pandemic period, such as dedicated laboratories. Of course, today, we can add a third dedicated method: one based on video conferencing with screen sharing. based on a slide viewer or scope system with a webcam and a network transmitter to tablet/smartphone, even when not present (Figure 4B), such as, for example, the DMshare system (Leica Microsystems Co., Nussloch GmbH, Germany). Both have allowed to free up important resources in this pandemic period, such as dedicated laboratories. Of course, today, we can add a third dedicated method: one based on video conferencing with screen

*Healthcare* **2021**, *9*, x FOR PEER REVIEW 6 of 13

**Figure 4.** The two different types of training: (**A**) using a very large tablet; (**B**) using the DMSHARE. **Figure 4.** The two different types of training: (**A**) using a very large tablet; (**B**) using the DMSHARE.

#### *3.5. The Standardization: A Slower Standardization Rate When Compared to Digital Radiology 3.5. The Standardization: A Slower Standardization Rate When Compared to Digital Radiology*

The standardization of imaging in PD has had and is having a more tortuous road than digital radiology, wherein, thanks to DICOM, since the 1990s [11], a rapid process of digitization and compatibility of the diagnostic tools of the organs and functions has been initiated (echo, NMR, CT, PET, etc.). The standardization of imaging in PD has had and is having a more tortuous road than digital radiology, wherein, thanks to DICOM, since the 1990s [11], a rapid process of digitization and compatibility of the diagnostic tools of the organs and functions has been initiated (echo, NMR, CT, PET, etc.).

Standardization in this area started with a slower process, and consequently, the compatibility between different manufacturers towards the standard has been delayed Standardization in this area started with a slower process, and consequently, the compatibility between different manufacturers towards the standard has been delayed [6]. Today, DICOM WSI http://dicom.nema.org/Dicom/DICOMWSI/, accessed on

Today, DICOM WSI http://dicom.nema.org/Dicom/DICOMWSI/ [12] is used as 6 July 2021, [12] is used as standard in DP.

These images are exceptionally large.

This standard considers the *whole slide images* (WSI)s in DP. This standard considers the *whole slide images* (WSI)s in DP. These images are exceptionally large.

As described in [12], a typical sample may be 20 mm × 15 mm in size and may be digitized with a resolution of 0.25 micrometers/pixel (conventionally described as *microns per pixel*, or *mpp*); here in the following we recall the characteristics reported in [12].

Most optical microscopes have an eyepiece which provides 10× magnification, so using a 40× objective lens results in 400× magnification.

Although instruments that digitize microscope slides do not use an eyepiece and may not use microscope objective lenses, by convention, images captured with a resolution of 0.25 mpp are referred to as 40×, images captured with a resolution of 0.5 mpp are referred to as 20X, etc.

The resulting image is therefore about 80,000 × 60,000 pixels, or 4.8 Gp.

Images are usually captured with 24-bit color, so the image data size is about 15GB.

This is a typical example, but larger images may be captured. Sample sizes up to 50 mm × 25 mm may be captured from conventional 1 × 3 slides, and even larger samples may exist on 2 × 3 slides.

Images may be digitized at resolutions higher than 0.25 mpp; some scanning instruments now support oil immersion lenses, which can magnify up to 100×, yielding 0.1 mpp resolution. Some operations described in [12] may further enlarge the data occupancy [12].

For example, a sample of 50 mm × 25 mm could be captured at 0.1 mpp with 10 Zplanes in the Z-stack, yielding a stack of 10 images of dimension 500,000 × 250,000 pixels. *Each plane would contain 125 Gp, or 375 GB of data, and the entire image dataset would contain a staggering 3.75 TB of data.*

#### **4. Towards the Revolution of the Digital Pathology and Artificial Intelligence**

*4.1. What Is Emerging in the Application of the Artificial Intelligence in Digital Pathology*

We carried out research with the aim of identifying the work to be completed, in terms of challenges and opportunities, towards stabilizing the use of artificial intelligence in DP, and then integrating what is highlighted with the considerations on digital pathology that we carried out in the previous section.

A quick look at PubMed with the following search key:

(*digital pathology* [Title]) AND (*artificial intelligence* [Title]) currently reports 17 works [6,13–28].

Among these works, one respects the search rule:

(digital pathology [Title]) AND (artificial intelligence [Title]) AND (COVID-19) [20] that is, it relates to COVID-19.

What is highlighted by these works (many of which are editorial and/or opinion) from a general point of view are the following aspects. The first aspect is that when scholars talk about artificial intelligence in digital pathology, they refer more to the aspects of imaging and essentially histological imaging. The second aspect is that scholars begin to identify interesting perspectives—for example, in oncology [15,25] or in toxicology [14,24]. The last aspect, in line with our objective, is that scholars are interrogating the work to be completed in a prospective way [25–28].

Important perspectives have been identified for example:


In the review that we have preselected as the only study linked to COVID-19 [20], it is highlighted that the effects of COVID-19 on research and clinical trials have also been significant with changes to protocols, suspensions of studies and the redeployment of resources to COVID-19 also useful for the applications of AI in DP. In this article, the authors explore the specific impact of COVID-19 on clinical and academic pathology and explore how digital pathology and artificial intelligence can play a key role in safeguarding clinical services and pathology-based research in the current climate and in the future.

We have identified *four prospective studies* that identify the critical issues and the work to be carried out [25–28].

*The first study*, although [25] it is not a review but an opinion, clearly identifies and discusses the critical issues in precision oncology by identifying some points on which to focus attention. The study aimed to provide a broad framework for incorporating AI and machine learning tools into clinical oncology, with an emphasis on biomarker development. *They discussed some of the challenges related to the use of AI, including the need for well-curated validation datasets, regulatory approval, and fair reimbursement strategies.*

*The second study* is an interesting review on the critical issues and the work still to be completed to arrive at the clinical routine [26]. This work highlights that while this is an exciting development that could discover novel predictive clinical information and potentially address international pathology workforce shortages, there is a clear need for a robust and evidence-based framework in which to develop these new tools in a collaborative manner that meets regulatory approval. *With these issues in mind, they have set out a roadmap to help academia, industry, and clinicians develop new software tools to the point of approved clinical use.*

The third study is an interesting review [27] that highlights that the advent of wholeslide imaging (WSI), the availability of faster networks, and cheaper storage solutions have made it easier for pathologists to manage digital slide images and share them for clinical use. In parallel, unprecedented advances in machine learning have enabled the synergy of artificial intelligence and digital pathology, which offers image-based diagnosis possibilities that were once limited only to radiology and cardiology. Integration of digital slides into the pathology workflow, advanced algorithms, and computer-aided diagnostic techniques extend the frontiers of the pathologist's view beyond a microscopic slide and enable true utilization and integration of knowledge in new manner; therefore, it is important to focus on the WSI, now standardized in DICOM WSI and as radiologists and cardiologists move in line with the standards.

*The fourth study* is an interesting review [28] where the authors provide a realistic account of all the challenges of adopting AI algorithms in digital pathology from both engineering and pathology perspectives.

In the work, we found an interesting and shareable outline of the challenges of AI in digital pathology that naturally recalls what emerges in the other three interesting prospective studies [25–28] and lends itself well to the objectives of our study.

#### *4.2. What Are the Perfectives and the Work to Be Carried out to Fully Integrate Artificial Intelligence in Digital Pathology?*

#### 4.2.1. The Guiding Approach

In Section 3, we highlighted the characteristics and criticalities of the digital pathology on which the AI will have to rely and, in particular, which ones will have to be taken into account in routine applications.

We have, furthermore, seen above that to make AI a consolidated reality in digital pathology, it is necessary: (a) proceed with standardization processes including the need for well-curated validation datasets, regulatory approval and fair reimbursement strategies [25], (b) define roadmaps to help academia, industry, and clinicians to develop new software tools to the point of approved clinical use through concerted actions [26], (c) focus on the WSI, now standardized in DICOM WSI and, as the radiologists and cardiologists move in line with the DICOM standards [27], (d) provide a realistic account of all chal-

lenges of adopting AI algorithms in digital pathology from both engineering and pathology perspectives [28].

#### 4.2.2. Future Challenges

In their exhaustive review, Hamid Reza Tizhoosh and Liron Pantanowitz [28] recently categorized the challenges to be faced and also the evident opportunities. We fully share this useful approach organized as a useful grid. We summarize this briefly, referring to the review for an in-depth view.

#### Challenges in AI in Digital Pathology

The challenges that digital pathology presents for the integration of AI have been identified in [28]'s 10 *challenges* (Figure 5, table in the left)*:*

#### 1. *Lack of labeled data*

The AI algorithms require a large set of good-quality training images. These training images must ideally be "labeled" (i.e., annotated). This is not easily feasible in DP.

#### 2. *Pervasive variability*

There are several basic types of tissue (e.g., epithelium, connective tissue, nervous tissue, and muscle). The actual number of patterns derived from these tissues from a computational point of view is nearly infinite if the histopathology images are to be "understood" by computer algorithms.

#### 3. *Non-Boolean nature of diagnostic tasks*

In pathology, not all can be summarized into two possible values such as "yes" or "no" (e.g., benign, or malignant). This is a too drastic a simplification of the complex nature of the diagnosis in this field. However, today, this is really not an issue; indeed, discrete variables (e.g., 1, 2, 3, 4) can be managed by machine learning, and there are also available methods based on regression machine learning for continuous variables, as reported in [29], for example.

#### 4. *Dimensionality obstacle*

As we have highlighted in Section 3, the WSI deals with gigapixel digital images of extremely large dimensions up to 3.75 TB. Deep ANNs used in AI act on much smaller image dimensions (i.e., not larger than 350 by 350 pixels).

#### 5. *Turing test dilemma*

The pathologist has the last word on the decision process when AI solutions are integrated in the workflow. Thus, full automation is probably neither possible, it seems, nor wise, as the Turing test postulates.

#### 6. *Uni-task orientation of weak artificial intelligence*

What we consider today is mostly "weak AI" in contrast with strong AI, also called artificial general intelligence (AGI). Deep ANNs belong to the class of weak AI algorithms, as they are designed to perform only one task. Therefore, we need to separately train multiple AI solutions for different tasks. This obviously has implications.

#### 7. *Affordability of required computational expenses*

Solutions with AI use graphical processing units (GPUs), highly specialized electronic circuits for fast processing of pixel-based data (i.e., digital images and graphics). These devices are expensive, and their adoption needs specific financial programs.

#### 8. *Adversarial attacks—The noise in the deep decisions*

This is a common problem in AI; a little change in a pixel, for example, due to the noise may cause a completely different output in the ANN.

#### 9. *Lack of transparency and interoperability*

The major drawbacks of artificial neural networks (ANN)s when used as classifiers is the lack of interoperability and transparency. Some consider ANNs to enclose a "black box" after the training. The major drawbacks of artificial neural networks (ANN)s when used as classifiers is the lack of interoperability and transparency. Some consider ANNs to enclose a "black box" after the training.

#### 10. *Realism of artificial intelligence* 10. *Realism of artificial intelligence*

There is currently optimism about the opportunities of ANNs, as has been highlighted above in the studies [13–28]. There are several difficulties with deploying AI tools in practice depending on the expectance and the objectives of the pathologist. There is no doubt that three are the preliminary requirements to improve this: (1) ease of use, (2) financial return on investment connected to the application, and (3) trust (such as, for example, the accountable performances). There is currently optimism about the opportunities of ANNs, as has been highlighted above in the studies [13–28]. There are several difficulties with deploying AI tools in practice depending on the expectance and the objectives of the pathologist. There is no doubt that three are the preliminary requirements to improve this: (1) ease of use, (2) financial return on investment connected to the application, and (3) trust (such as, for example, the accountable performances).

**Figure 5.** The challenges and cross-cutting issues emerging in the application of AI in DP. **Figure 5.** The challenges and cross-cutting issues emerging in the application of AI in DP.

Further Cross-Cutting Issues Further Cross-Cutting Issues

We agree with the categorization identified by Hamid Reza Tizhoosh and Liron Pantanowitz [28], and I believe that it can be used as a reference for evaluating the future efforts of AI in digital pathology. Without introducing new challenges in detail, we would like to integrate the analysis with what emerged in Section 3 and in the other three selected prospective studies discussed above [25–28]. We agree with the categorization identified by Hamid Reza Tizhoosh and Liron Pantanowitz [28], and I believe that it can be used as a reference for evaluating the future efforts of AI in digital pathology. Without introducing new challenges in detail, we would like to integrate the analysis with what emerged in Section 3 and in the other three selected prospective studies discussed above [25–28].

There are in fact aspects to be highlighted that act in a transversal way and are decisive for facing the 10 challenges identified in the categorization. There are in fact aspects to be highlighted that act in a transversal way and are decisive for facing the 10 challenges identified in the categorization.

*Cross-cutting issues to be considered in the challenges* (Figure 5, table in the right). *Cross-cutting issues to be considered in the challenges* (Figure 5, table in the right).


#### **5. Conclusions and Work in Progress**

#### *5.1. The Evidences in the Study*

In this study, the introduction of artificial intelligence in digital pathology was addressed. The study first tackled the second revolution in diagnostic pathology determined by the introduction of digital pathology techniques [1–6]. There is no doubt that most of the applications of AI take place in diagnostic imaging and that, therefore, AI rests on the imaging techniques used in digital pathology.

In analysing the important aspects of digital pathology, some important points/steps were noted:


We then questioned the state of the next revolution that is anticipated due to the introduction of AI in DP. Through an overview of some important studies, some important development guidelines have been identified and, in line with the objectives of this study, the challenges to be addressed in detail and the transversal problems as they emerge both from the overview and from the characteristics and problems of digital pathology highlighted in the section dedicated to this discipline. The *10 challenges* were therefore recalled, starting from the grid identified in [28], and eight emerged transversal issues to be considered in these challenges were introduced and discussed (Figure 5).

#### *5.2. Actual Developments and Future Work*

All that is highlighted in the *cross-cutting issues* is, in a certain sense, of strong scientific interest and needs attention if we think of a routine introduction of AI in digital pathology. A point where we intend to contribute is that (no. 7) relating to acceptance based on surveys on key figures (no. 8), which is preparatory to standardization actions (no. 6 and no. 3–4). Inheriting the experience gained from previous studies [1,5], in which we had developed paper surveys for this purpose (relating to the introduction of digital pathology first in eHealth and then in mHealth), we are developing an electronic survey as a tool to be used with this purpose and we are using it to investigate this.

#### *5.3. Limitations of the Study*

The overview in the section was conducted with the search key "Artificial Intelligence", wanting to stay on a higher and general level regarding the topic in line with the objectives of the study. Other more specific searches can be executed on aspects of a lower hierarchical level such as those relating to the algorithms of use. Artificial intelligence uses a myriad of different methodologies, techniques, and approaches that deserve specific review and research extended to non-medical databases, even if we are dealing with medical problems.

A long discussion deserves a targeted approach in the collection of *medical knowledge* in this area relating to supervised ANNs and unsupervised ANNs to collect successful and/or unsuccessful experiences.

A key search, for example, limited to the medical database PubMed of (digital pathology [Title]) AND (deep learning [Title]) led, at the date of this study, to 14 results [30], of which one was included in the one we made.

Another example of research on the same database of (digital pathology [Title]) AND (machine learning [Title]) led, at the date of this study, to seven results [31], of which four were included in the one we made above.

Such a research is more closely related to the specific performance of algorithms in DP and can highlight important development opportunities that must certainly be taken into account in any wide-ranging reviews.

**Author Contributions:** Conceptualization, D.G. and M.R.G.; methodology, D.G.; software, D.G.; validation, D.G. and M.R.G.; formal analysis, All; investigation, All; resources, All; data curation, D.G.; writing—original draft preparation, D.G.; writing—review and editing, All; visualization, All; supervision, All; project administration, All; funding acquisition, none. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** Not applicable.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


**Maria Rosaria Giovagnoli <sup>1</sup> , Sara Ciucciarelli <sup>1</sup> , Livia Castrichella <sup>1</sup> and Daniele Giansanti 2,\***


**Abstract:** *Motivation:* This study deals with the introduction of artificial intelligence (AI) in digital pathology (DP). The study starts from the highlights of a companion paper. *Objective:* The aim was to investigate the consensus and acceptance of the insiders on this issue. *Procedure:* An electronic survey based on the standardized package Microsoft Forms (Microsoft, Redmond, WA, USA) was proposed to a sample of biomedical laboratory technicians (149 admitted in the study, 76 males, 73 females, mean age 44.2 years). *Results:* The survey showed no criticality. It highlighted (a) the good perception of the basic training on both groups, and (b) a uniformly low perceived knowledge of AI (as arisen from the graded questions). Expectations, perceived general impact, perceived changes in the *work-flow*, and worries clearly emerged in the study. *Conclusions:* The of AI in DP is an unstoppable process, as well as the increase of the digitalization in the *health domain*. Stakeholders must not look with suspicion towards AI, which can represent an important resource, but should invest in monitoring and consensus training initiatives based also on electronic surveys.

**Keywords:** e-health; medical devices; m-health; digital-pathology; picture archive and communication system; artificial intelligence; cytology; histology; diagnostic pathology

#### **1. Introduction**

In a complementary study [1] we dealt with the introduction of artificial intelligence (AI) in digital pathology (DP). This could lead to a *second revolution* in pathological diagnostics (starting from the *first revolution* determined by the introduction of DP techniques both in *eHealth* and *mHealth* [2,3]). Most AI applications [1] take place in diagnostic imaging. However, there are many important implications related to the introduction of AI. These implications involve *other disciplines* (not only connected to imaging) and *other activities*, from the *work-flow* to the training. In our study [1] we recalled the passages that led to the first revolution of diagnostic pathology, represented by DP. We dedicated particular attention to the critical issues, given that AI will rely heavily on it. In the same study, we highlighted the opportunities and the challenges of AI according to the most recent studies [4–20]. Some important development guidelines have been identified. The DP developments with AI have been identified [20]. AI shows in DP (A) the potentiality to access and correlate large amount of data, and (B) direct prospective in the world of diagnostics.

Regarding *A*, both radiological and pathology images are stored in the *picture archiving and communication systems* (PACs). Moreover, with the introduction of electronic health records (EHRs), systematic collections of patient health information have been made available. They include qualitative data, medical records, and laboratory and diagnostics information. AI, if applied to these large digital stores, could prove useful for epidemiological, clinical, and research studies.

Regarding *B*, two aspects are emerging:

**Citation:** Giovagnoli, M.R.; Ciucciarelli, S.; Castrichella, L.; Giansanti, D. Artificial Intelligence in Digital Pathology: What Is the Future? *Part 2: An Investigation on the Insiders*. *Healthcare* **2021**, *9*, 1347. https://doi.org/10.3390/ healthcare9101347

Academic Editors: Tin-Chih Toly Chen and Mariano Cingolani

Received: 15 July 2021 Accepted: 9 October 2021 Published: 11 October 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

	- The pathological diagnosis of the future could merge proteomics and genomics in the BIG-DATA. • The pathological diagnosis of the future could merge proteomics and genomics in the BIG-DATA. •development of the DP, due to the introduction of *whole-slide scanners* and the *progress of computer vision algorithms*, have significantly grown the usage of AI. It can Regarding *B*, two aspects are emerging: • The development of the DP, due to the introduction of *whole-slide scanners* and the

The challenges to tackle and the evident opportunities of AI in DP were recently categorized in [19]. These challenges were therefore recalled in [1], starting from the grid identified in [19]. The following transversal issues to be considered in these challenges were introduced and discussed [1]: The challenges to tackle and the evident opportunities of AI in DP were recently categorized in [19]. These challenges were therefore recalled in [1], starting from the grid identified in [19]. The following transversal issues to be considered in these challenges were introduced and discussed [1]: perform tumor diagnosis, subtyping, grading, staging, and prognostic prediction • The pathological diagnosis of the future could merge proteomics and genomics in the BIG-DATA. The challenges to tackle and the evident opportunities of AI in DP were recently cat*progress of computer vision algorithms*, have significantly grown the usage of AI. It can perform tumor diagnosis, subtyping, grading, staging, and prognostic prediction • The pathological diagnosis of the future could merge proteomics and genomics in the BIG-DATA.


*Healthcare* **2021**, *9*, x 2 of 9

*Healthcare* **2021**, *9*, x 2 of 9

Regarding *B*, two aspects are emerging:


All that is highlighted in the above cross-cutting issues is of strong scientific interest. These issues are basic to plan a routine introduction of AI in DP. All that is highlighted in the above cross-cutting issues is of strong scientific interest. 6. Need for standardization actions. 7. Extensive acceptance surveys on professionals. 5.New training models must adapt to AI in DP. 6.Need for standardization actions.

We intend with this study to concentrate on some of the points detected. These issues are basic to plan a routine introduction of AI in DP. We intend with this study to concentrate on some of the points detected. 8. Need to focus on all the professionals involved. 7.Extensive acceptance surveys on professionals.

We intend to propose a survey (*point 7*) focused on the professionals involved (*point 8*) to investigate the state of acceptance and the consensus on the introduction of AI in DP. Prior to this study, the experience reported in [21] focused on pathological diagnostics (*a single aspect of DP*), on a *single profession,* and proposed a non-validated and non-standardized questionnaire on the acceptance of AI in general. Despite limitations, several interesting findings were uncovered. Overall, respondents carried generally positive attitudes towards AI, excitement in AI as a diagnostic tool to facilitate improvements in *work-flow* efficiency, and quality assurance in pathology. Importantly, even within the most optimistic cohort, a significant number of respondents endorsed concerns about AI, including the potential for job displacement and replacement. Overall, around 80% of respondents predicted the introduction of AI technology in the pathology laboratory within the coming decade. The study focused on one single professional [21]; however, many other professionals are revolving around the introduction of AI in DP, ranging from the pathologist up to the biomedical laboratory technician. We intend to propose a survey (*point 7*) focused on the professionals involved (*point 8*) to investigate the state of acceptance and the consensus on the introduction of AI in DP. Prior to this study, the experience reported in [21] focused on pathological diagnostics (*a single aspect of DP*), on a *single profession,* and proposed a non-validated and non-standardized questionnaire on the acceptance of AI in general. Despite limitations, several interesting findings were uncovered. Overall, respondents carried generally positive attitudes towards AI, excitement in AI as a diagnostic tool to facilitate improvements in *work-flow* efficiency, and quality assurance in pathology. Importantly, even within the most optimistic cohort, a significant number of respondents endorsed concerns about AI, including the potential for job displacement and replacement. Overall, around 80% of respondents predicted the introduction of AI technology in the pathology laboratory within the coming decade. The study focused on one single professional [21]; however, many other professionals are revolving around the introduction of AI in DP, ranging from the pathologist up to the biomedical laboratory technician. All that is highlighted in the above cross-cutting issues is of strong scientific interest. These issues are basic to plan a routine introduction of AI in DP. We intend with this study to concentrate on some of the points detected. We intend to propose a survey (*point 7*) focused on the professionals involved (*point 8*) to investigate the state of acceptance and the consensus on the introduction of AI in DP. Prior to this study, the experience reported in [21] focused on pathological diagnostics (*a single aspect of DP*), on a *single profession,* and proposed a non-validated and non-standardized questionnaire on the acceptance of AI in general. Despite limitations, several interesting findings were uncovered. Overall, respondents carried generally positive attitudes towards AI, excitement in AI as a diagnostic tool to facilitate improvements in *work-flow* efficiency, and quality assurance in pathology. Importantly, even within the most optimistic cohort, a significant number of respondents endorsed concerns about AI, including the potential for job displacement and replacement. Overall, around 80% of respondents predicted the introduction of AI technology in the pathology laboratory within the com-8. Need to focus on all the professionals involved. All that is highlighted in the above cross-cutting issues is of strong scientific interest. These issues are basic to plan a routine introduction of AI in DP. with this study to concentrate on some of the points detected. We intend to propose a survey (*point 7*) focused on the professionals involved (*point 8*) to investigate the state of acceptance and the consensus on the introduction of AI in DP. Prior to this study, the experience reported in [21] focused on pathological diagnostics (*<sup>a</sup> single aspect of DP*), on a *single profession,* and proposed a non-validated and non-standardized questionnaire on the acceptance of AI in general. Despite limitations, several interesting findings were uncovered. Overall, respondents carried generally positive attitudes towards AI, excitement in AI as a diagnostic tool to facilitate improvements in *work-flow* efficiency, and quality assurance in pathology. Importantly, even within the most optimistic cohort, a significant number of respondents endorsed concerns about AI, including the potential for job displacement and replacement. Overall, around 80% of respondents

There are many other aspects to be taken into consideration besides the diagnostic aspects [21]. We must consider, for example [1,19,20], the peculiarity of digital cytology and of digital histology, omics (e.g., genomics and proteomics), integration with BIGDATA, integration with historical and clinical data of the patient, the search for slide labelling, quality control, the integration of DP with digital radiology, training, risk analysis, therapy, and prevention. There are many other aspects to be taken into consideration besides the diagnostic aspects [21]. We must consider, for example [1,19,20], the peculiarity of digital cytology and of digital histology, omics (e.g., genomics and proteomics), integration with BIG-DATA, integration with historical and clinical data of the patient, the search for slide labelling, quality control, the integration of DP with digital radiology, training, risk analysis, therapy, and prevention. ing decade. The study focused on one single professional [21]; however, many other professionals are revolving around the introduction of AI in DP, ranging from the pathologist up to the biomedical laboratory technician. There are many other aspects to be taken into consideration besides the diagnostic aspects [21]. We must consider, for example [1,19,20], the peculiarity of digital cytology and of digital histology, omics (e.g., genomics and proteomics), integration with BIGpredicted the introduction of AI technology in the pathology laboratory within the coming decade. The study focused on one single professional [21]; however, many other professionals are revolving around the introduction of AI in DP, ranging from the pathologist up to the biomedical laboratory technician. There are many other aspects to be taken into consideration besides the diagnostic aspects [21]. We must consider, for example [1,19,20], the peculiarity of digital cytology

The goal of our study was to The goal of our study was to DATA, integration with historical and clinical data of the patient, the search for slide laand of digital histology, omics (e.g., genomics and proteomics), integration with BIG-


#### Analyze the outcome. tirety [1,19,20] (not limited to pathological diagnostics) and the involved professionals. to the introduction of AI in DP, considering both the opportunities of AI in their entirety [1,19,20] (not limited to pathological diagnostics) and the involved profession-**2. Materials and Methods**

 Submit it electronically to a first sample of insiders. Analyze the outcome. als. Submit it electronically to a first sample of insiders. Analyze the outcome. In line with the aim of the study, we decided to propose a survey to investigate the acceptance and the consensus of the insiders. Preliminarily, we addressed the aspects of privacy and data security. The questionnaire was checked for the compliance to the European GDPR 679/2016 and the Italian Decree 101/2018, as required by the Data Protection Offices. The questionnaire was planned as anonymous. The topic did not concern clinical

trials on humans, but only opinions and expressions of their thoughts. In consideration of this, it was not considered necessary to proceed with the formal approval procedures from the Institutional Review Board (see footnote at the end). The standard Microsoft Forms package (Microsoft Forms, Redmond, WA, USA) was chosen.

This package is also available with a free Microsoft account (live, outlook, or hotmail, for example), but in this case, it has important limitations (for example, the maximum number of participants is limited to 200). The data acquired by means of Microsoft Forms represent a public register from a legal point of view. Therefore, data need to be strongly protected by means of a strong cybersecurity approach. This is not feasible using only a free Microsoft account.

Companies that have centrally installed the Microsoft 365 App Business Premium suite have Microsoft Forms available to their users with greater potential than the free version (for example, the maximum limit of participants is raised to 50,000). All users can have access through their own domain account guaranteed by the corporate cybersecurity standards (which must comply with the international regulations in force) supported by network and system security tools and policies managed by the company. Specific checks are possible on the IPs (registering, for example, the duplicate access for further dataprocess). Data are therefore protected by the corporate cybersecurity systems, guaranteeing (at least from the system point of view) the inviolability of the data. In consideration of this, we have decided to use the software Microsoft Forms, provided through the Microsoft 365 Business Premium suite, to design an electronic survey. It is the tool recommended by the company's DPO. It should be noted that if a tool other than those available in this suite (e.g., Google forms or Survey Monkey) had been used, the DPO would have requested a specific report and a cybersecurity audit. The authorization to use it would not have been guaranteed. The use of both an internally recommended tool (respecting the cybersecurity) and the plan to submit the electronic survey (eS) anonymously simplified the authorization process. However, we decided to maintain the database as a register, respecting the security criteria identified by the company rules in accordance with the law. The procedure used in the design and submission of the survey adhered to the *SURGE Checklist* [22].

We decided to submit the survey to the key professionals and therefore disseminated it through social media, such as Facebook, LinkedIn, Twitter, Instagram, WhatsApp, Association Sites, and in general, following a *peer-to-peer* dissemination. We submitted the survey to biomedical laboratory technicians during their course of study (*BLT-DCS*) and after the course of the study (*BLT-ACS*). The interactive survey is available in [23]. A print can also be found in [24].

Two questions (N.2 and N.3) stratify by age and sex [23,24]. Two initial questions (N.4 and N.5) categorize the sample on the basis of the training background. In consideration of the objective of this study and the survey, we also managed the survey as a virtual focus group, with careful considerations to the consensus issues related to all the aspects of the introduction of AI in DP [1,19,20]. We started from the training up to the relationships and integration with omics, BIG-DATA, and digital radiology. The methodological approach primarily involves submitting both to *BLT-DCS* and *BLT-ACS* surveys. Figure 1 shows the CONSORT diagram. The final records were 211 in number. Two records were excluded because the answers to the open questions were not coherent.

The subjects passing the requirements for the inclusion according to the selection criteria (*BLT-DCS* or *BLT-ACS*) were 149 (Table 1).

The quantitative variables depended on subjective answers based on qualitative perceptions (see for example in the following the graded questions or the modules in the Likert scale). The survey used *open question, choice question, multiple choice questions, Likert questions,* and *graded questions*.

We established a six-level psychometric scale for the *Likert scale* and the *graded questions*. It was possible therefore to assign a minimum score of one and a maximum of six with a *theoretical mean value* (TMV) of 3.5. We can refer to the TMV for comparison in the analysis of the answers. An average value of the answers below TMV indicates a more negative

than positive response. An average value above TMV indicates a more positive than negative response. *Healthcare* **2021**, *9*, x 4 of 9

**Figure 1.** The CONSORT diagram. **Figure 1.** The CONSORT diagram.

**Table 1.** Characteristics of the admitted to the study the DCS and the ACS.

**Table 1.** Characteristics of the admitted to the study the DCS and the ACS.


The quantitative variables depended on subjective answers based on qualitative perceptions (see for example in the following the graded questions or the modules in the Likert scale). The survey used *open question, choice question, multiple choice questions, Likert questions,* and *graded question*s. The trend of each one of these variables, estimated by an average value, can move in both the two directions, toward the higher score of 6 or toward the lower score of 1, suggesting for a two-tailed test. For the variables related to the *multiple-choice* questions, we planned a frequency analysis.

We established a six-level psychometric scale for the *Likert scale* and the *graded questions*. It was possible therefore to assign a minimum score of one and a maximum of six For the verification of data normality, we used the Shapiro–Wilk test that is preferable for small samples such as ours.

with a *theoretical mean value* (TMV) of 3.5. We can refer to the TMV for comparison in the analysis of the answers. An average value of the answers below TMV indicates a more We applied Student's *t*-test (with a *p*-value <0.01 for the significance of the difference), when comparing the values between the two groups.

negative than positive response. An average value above TMV indicates a more positive than negative response. We applied the χ 2 test (with a *p*-value <0.01 for the significance) in the frequency analysis. The software SPSS Statistics version V.24 was used in the study.

The trend of each one of these variables, estimated by an average value, can move in both the two directions, toward the higher score of 6 or toward the lower score of 1, sug-The Cohen's d effect size was estimated to be 0.498. Samples with N > 60 were estimated suitable to the study.

gesting for a two-tailed test. For the variables related to the *multiple-choice* questions, we We established a six-level psychometric scale in the graded questions and in the Likert scale.

We applied Student's *t*-test (with a *p-*value <0.01 for the significance of the differ-

planned a frequency analysis. For the verification of data normality, we used the Shapiro–Wilk test that is prefera-The survey was proposed from 1 June 2021 until 23 August 2021.

analysis. The software SPSS Statistics version V.24 was used in the study.

ence), when comparing the values between the two groups.

ble for small samples such as ours.

#### **3. Results**

Table 2 shows the answers to the graded *questions*. Both questions Q6 and Q7, not focused directly on AI, received an average response value above the TMV threshold.


**Table 2.** Answers for the graded questions.

However, Q6 showed a significantly higher value in the student group (*p*-value < 0.01), while Q7 showed a consistent value between the two groups (*p*-value = 0.134 >> 0.01).

The responses related to AI, Q8–10, showed a value below the current TMV threshold in the two groups (*p*-value < 0.01).

Tables 3 and 4 highlight the outcomes for the two Likert scales in detail. In the first Likert scale (Table 3), *imaging* (cytological and histological) received the highest score for the two groups, followed by applications in *omics and quality control*. Table 4 shows the significant highest values for the first group in the second Likert scale dedicated to other sectors of applications.

**Table 3.** Detailed answers in the Likert scale to the question of "In which specific sectors of biomedical diagnostics do you think the introduction of artificial intelligence is most promising?".



**Table 4.** Detailed answers in the Likert scale to the question of "In which more general sectors do you think artificial intelligence is useful?".

The *multiple-choice questions* are useful for obtaining strategic information, for example, for scientific societies or consensus activities. We decided to proceed as follows, in consideration of the peculiarity of these modules. We analyzed the two samples joined into one sample and performed a statistical approach based on a frequency analysis, using the test described in the methods.

For question Q13 "*I think artificial intelligence in my field*", the two most popular statements were "*It will be useful but complementary*" number of votes = *83* and "*It will not catch on*" number of votes = *78* (*p*-value = 0.008).

For question Q14 "*How can I be of use to AI in my filed*", the two most popular statements were "*In performance monitoring"* number of votes = *90* and "*As an operational manager of its use*" number of votes = *81* (*p*-value = 0.008).

For question Q15 "*How will AI help me*", the two most popular statements were "*Increased automatism*" number of votes = *79* and "*Reduction of physical fatigue*" number of votes = *61* (*p*-value = 0.009).

#### **4. Discussion and Conclusions**

The use of AI is increasingly spreading in many medical sectors.

A particularly important area for applications is that of images. A simple search on PubMed with the key

*(artificial intelligence [Title/Abstract]) AND (image [Title/Abstract])*

shows 2290 results as of 23 August 2021 (907 in 2021).

*This justifies the need of focusing on studies of acceptance, in consideration of both the interest of the scholars and the possible opportunities in the clinical routine*.

Some studies are also demonstrating the importance of AI tools, not only in imaging, but also in other applications where *data mining from large volumes of data must be applied*.

For example, the study reported in [25] showed how AI is useful for determining cardiovascular risk in athletes through *data mining of distributed databases*.

The COVID-19 pandemic has also highlighted *the broad-spectrum potential of AI*. In a recent review [26], for example, relevant papers were selected that address the adoption of artificial intelligence and new technologies in the management of pandemics and communicable diseases such as SARS-CoV-2.

These studies focused on environmental measures; acquisition and sharing of knowledge in the general population and among clinicians; development and management of drugs and vaccines; remote psychological support of patients; remote monitoring, diagnosis, and follow-up; and maximization and rationalization of human and material resources in the hospital environment. The study described in [27] showed that *AI-based scores with a purely data-driven selection of features* are feasible and effective for the prediction of mortality among patients with COVID-19 pneumonia.

*The three illustrated potentials [25–27] are also important in DP. In fact, in DP, the need for categorizing images merges with the need to make decisions and/or deduce approaches through actions on large databases and data sets or with other needs not based on medical images [1,19,20]. The implications are multifaceted. It is necessary to carry out direct studies on the opinion of insiders in view of the introduction of the clinical routine of AI*. *Therefore, the need and the justification of studies such as ours that tackle the introduction of AI focusing on acceptance and with a broad approach clearly emerges from these articles [25–27]*.

Very few studies have begun to address the insiders' opinion on the introduction of AI in DP. By searching in PubMed with the key

*((digital pathology [Title/Abstract]) AND (artificial intelligence [Title/Abstract])) AND (survey [Title/Abstract])-even with alternative terms to the survey-*

we found as of 23 August 2021 only two studies based on non-validated and nonstandardized questionnaires.

The first study [28] was conducted at a scientific meeting (the 14th Banff Conference). Since the meeting, a survey with international participation of mostly pathologists (81%) was conducted, showing that whole slide imaging is available at the majority of centers (71%), but that artificial intelligence (AI)/machine learning was only used in ≈12% of centers, with a wide variety of programs/algorithms employed.

The second study [29] reports the results of the Japanese questionnaire survey conducted in 2008–2009 on telepathology and virtual slide. Moreover, in addition to the questionnaire, the effectiveness of an experimental automatic pathology diagnostic aid system using computer artificial intelligence was investigated by checking its rate of correct diagnosis for given prostate carcinoma digital images.

This demonstrates the importance of focusing on wide-ranging survey studies in this field. From this research, it clearly emerges that specific studies, such as ours based on wide range questionnaires, have not been addressed until now. In fact, in the literature, there are currently only studies that deal with the topic only partially or secondarily [28,29].

This study was necessary to prepare a first survey dedicated to the acceptance of AI in DP focused on the insiders [1]. We submitted the survey on the professionals involved in the field. Many professionals are involved in the introduction of AI in DP, ranging from the bioengineer to the pathologist up to the biomedical laboratory technician. There is also no doubt that AI could represent a serious opportunity for the DP laboratories [5–18]. It is, however, the time to investigate the full introduction in the routine. The proposed study, for example, can be useful in view of consensus studies on the introduction of methods based on AI in DP in routine practices [1,19]. We have proposed a survey focused on these professionals that is, in an automatic manner, capable of electronically collecting their opinion and works as a structured virtual focus group.

The intent of this study was to carry out a first submission and to verify any criticalities in view of a wider use. There were no critical issues and the submission made it possible to collect information on a first sample of biomedical laboratory technicians in the training phase and subsequent phase.

A good perception of the basic training on both groups (albeit with a different score) and a uniformly low perceived knowledge of the use of AI emerged from the graded questions.

The *two Likert scales* made it possible to identify in a structured way, for the two groups, the wishes related to the use of AI in the medical field.

The *multiple-choice* questions, evaluated for the whole combined sample, allowed us to evaluate the perceived impact of AI in one's sector, the expectations towards AI, and the operational role towards AI. From a general point of view, the study presents three added values.

The *first added* value is [23,24] represented by the electronic tool with a wide range of aspects related to the use of AI in DP, having a direct impact on the *work-flow* and *job description* of the insiders.

The *second added* value is a contribution directed to respond to the need to tackle the challenges of the introduction of AI in DP. This product (after minimal changes) could be used by scientific and/or professional societies to monitor the evolution of the topic.

The *third added* value is represented by the outcome with reference to the two groups of *DCS* and *ACS* (promptly useful for the stakeholders).

From a general point of view, this article supports the initiatives that aim to facilitate the introduction of AI in a structured manner in DP. Future developments of the study foresee the enlargement of the submission to other professionals and a standardization for the scientific societies.

#### **5. Limitations**

This study represents a first step to investigate the acceptance and consensus on AI of insiders in the various applications and implications of DP. It was applied to a first professional and a first group of subjects. Future developments will have to include a broader submission involving other professionals, together with a review action by the scientific societies, in order to improve acceptance by the parties involved.

**Author Contributions:** Conceptualization, D.G. and M.R.G.; methodology, D.G. and L.C.; software, D.G. and S.C.; validation, D.G., M.R.G., and L.C.; formal analysis, all authors; investigation, all authors; resources, all authors; data curation, D.G.; writing—original draft preparation, D.G. and S.C.; writing—review and editing, D.G.; visualization, all authors; supervision, all authors; project administration, all authors; funding acquisition, none. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** Not applicable.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**

