i-SART: An Intelligent Assistant for Safety Analysis

in Radiation Therapy

Natalia Silvis-Cividjian

, Yijing Zhou

, Anastasia Sarchosoglou

and Evangelos Pappas

Vrije Universiteit Amsterdam, Department of Computer Science, Amsterdam, The Netherlands

University of West Attica, Department of Biomedical Sciences, Athens, Greece

Keywords: Radiation Therapy, Safety and Risk Management, Digital Assistive Technology, Failure Modes and Effect

Analysis (FMEA), Natural Language Processing (NLP), Generative AI, Synthetic Data.

Abstract: Along with surgery and chemotherapy, radiation therapy (RT) is a very effective method to treat cancer. The

process is safety-critical, involving complex machines, human operators and software. A proactive hazard

analysis to predict what can go wrong in the process is therefore crucial. Failure Modes and Effect Analysis

(FMEA) is one of the methods widely used for risk assessment in healthcare. Unfortunately, the available

resources and FMEA expertise strongly vary across different RT organizations worldwide. This paper

describes i-SART, an interactive web-application that aims to close the gap by bringing together best practices

in conducting a sound RT-FMEA. Central is a database that at present contains approximately 420 FMs

collected from existing risk assessments and cleaned from ambiguities and duplicates using NLP techniques.

Innovative is that the database is designed to grow, due to both user input and generative AI algorithms. This

is work in progress. First experiments demonstrated that using machine learning in building i-START is

beneficial. However, further efforts will be needed to search for better solutions that do not require human

judgment for validation. We expect to release soon a prototype of i-SART that hopefully will contribute to

the global implementation and promotion of safe RT practices.

1 INTRODUCTION

Cancer is the second leading cause of death

worldwide. About 40% of world’s population will be

diagnosed with cancer at some point during their

lifetimes (NCI, 2017). Radiation therapy (RT) is a

highly effective cancer management approach

received by approximately 50% of all patients. One

can say that RT is a field where healthcare meets

informatics. The process takes place in complex,

computer-controlled linear accelerators (linacs),

where high-energy ionizing radiation is used to

reduce or eliminate the tumor(s) and at the same time

sparing the healthy tissue (Fig. 1). The core RT team

consists of different healthcare professionals,

including radiation oncologists, medical physicists,

radiation therapists, dosimetrists and nurses. A

generic process RT process is illustrated in Fig. 1.

https://orcid.org/0009-0004-4668-5946

https://orcid.org/0009-0007-6487-0004

https://orcid.org/0009-0007-6487-0046

https://orcid.org/0000-0001-8052-6392

After a patient is referred for radiotherapy and

assessed by a radiation oncologist, the next step

involves an imaging exam, usually a CT localization

scan. On these images, the radiation oncologist

delineates the specific regions that have to be

irradiated and prescribes the dose in each of these

regions. After that, the treatment planning and

treatment delivery teams accurately follow this

prescription and deliver the needed radiation, by

using the linac and various types of software

products. During the whole process, Quality

Assurance (QA) and patient monitoring activities are

mandatory.

A few devastating accidents that occurred in the

last decades demonstrate that the RT process is

safety-critical - any mistake, be it caused by

hardware, software or humans, can have fatal

consequences (Leveson and Turner, 1993), (Borras et

420

Silvis-Cividjian, N., Zhou, Y., Sarchosoglou, A. and Pappas, E.

i-SART: An Intelligent Assistant for Safety Analysis in Radiation Therapy.

DOI: 10.5220/0012364400003657

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 2, pages 420-427

ISBN: 978-989-758-688-0; ISSN: 2184-4305

al., 2006). Therefore, RT is nowadays a strongly

regulated process, with safety standards in place

(Council of European Union, 2014), (IAEA, 2018).

According to these standards, an RT process needs to

be thoroughly assessed for all the risks it poses,

before obtaining permission to proceed (Huq et al.,

2016). This can be addressed with a proactive risk

analysis, which aims to anticipate failure modes (FM)

or hazards, defined as conditions that can lead to

incidents, or in other words, the various ways a

system can fail.

To conduct a safety assessment, the analysts can

choose from a range of systematic methods, such as

the traditional Failure Mode and Effects Analysis

(FMEA), Fault Tree Analysis (FTA) and Hazard and

Operability Analysis (HAZOP) (Pawlicki et al.,

2011), or the more modern Systems Theoretic

Accident Model and Process (STAMP) (Pawlicki et

al., 2016), (Silvis-Cividjian et al., 2020). All these

methods work in the same way: first, a team identifies

the potential hazards in a process, addresses their

causes and evaluates their effects, and finally

formulates appropriate mitigation measures.

Figure 1: a) The principle of RT; b) The geometry of RT,

where one can see that the tumor receives the highest

radiation dose (in red) and the healthy tissue the lowest (in

blue). From (Kane, 2014); c) A view from a typical

treatment facility room, where a radiation therapist needs to

distribute their attention over many computer screens.

Credits to A. Sarchosoglou; d) The workflow of a generic

RT process.

A general problem is that many RT departments

lack the time, training, or manpower required to

perform an in-depth risk assessment. Another problem

is that knowledge tends to remain compartmentalized

within departments, with safety analysis results often

not being shared widely. For example, currently there

is no centralized database with potential RT-specific

FMs that could be used as a reference by practitioners

who intend to conduct a proactive risk analysis. This

is a missed opportunity in our opinion, because

despite their diversity, all RT process workflows

feature in fact sufficient common FMs.

On the other hand, assistive and data mining

software applications, often powered by artificial

intelligence (AI), are rapidly emerging in all domains

of our daily life, including healthcare and RT.

Examples are software systems for electronic patient

dossiers, prediction of the response to a treatment,

disease risk assessment, or, specific for the RT

domain, radiation dose calculation, automatic

delineation of tumors and organs at risk on CT scan

images, or defacing of CT images of head-and-neck

cancer patients for privacy reasons, etc.

In this paper, we will present an attempt to close

the gap and improve RT safety worldwide with i-

SART, an online platform that assists practitioners in

performing an effective proactive FMEA-based

safety analysis. Central is a novel database that brings

together a large number of possible RT-specific FMs,

formulated in English and free of ambiguities or

duplicates. Innovative is that the database is designed

in such a way that new FM data can be fed not only

by safety-aware RT practitioners around the globe,

but also by state-of-the-art generative AI (GenAI)

algorithms. To the best of our knowledge, this is the

first attempt to use GenAI for synthetic FMs. As this

is work in progress, synthetic FMs were not included

yet in the i-SART database. Nevertheless, a prototype

of i-SART will be soon released for all interested RT-

practitioners.

The remainder of the paper is organized as

follows. In Section 2 we formulate the problem we try

to solve with i-SART, in Section 3 we will present the

design of i-SART, its database and user interface.

Section 4 will present some preliminary results and

Section 5 will outline our conclusions and future

work plans.

2 PROBLEM STATEMENT

First used by the US Military at the end of 40’s,

FMEA is a safety assessment method widely adopted

in systems engineering in 60’s (Arnzen, 1966).

i-SART: An Intelligent Assistant for Safety Analysis in Radiation Therapy

421

FMEA has been also widely used and recommended

for healthcare in general and RT in particular, in order

to prevent medical errors propagating and reaching

the patient (Ibanez, 2018), (Olch, 2014), (deRosier,

2002).

The general process flow of an FMEA is

illustrated in Fig. 2. The method is bottom-up,

meaning that for each component in the process, one

have to ask the question “How often would this

component fail, and what will happen if it fails?”. The

risk of each FM is evaluated using a Risk Priority

Number (RPN), calculated as the product of severity,

probability of occurrence, and detectability.

The result of an FMEA analysis is a list with all

possible FMs, ranked by their RPNs, their causes and

their effects, followed by measures to mitigate the

most critical ones. In an RT process, some FMs that

can occur are readily conceivable, such as “A wrong

patient is invited to the treatment room” or “The linac

gantry in rotation collides with the treatment couch”.

However, to conduct an analysis that will predict all

ways a process can fail is challenging. This task

demands considerable time, domain knowledge and

clinical experience. Unfortunately, these resources

may not always be readily available as the main task

of RT practitioners is to treat as many patients as

possible in a race against the clock with fast-evolving

cancer.

Figure 2: The workflow of an FMEA analysis.

To gain more insights into the current FMEA

practices within the RT, we recently conducted a

literature review that highlighted the diversity

between departments in the way they conduct their

FMEA (Sarchosoglou et al., 2022). A challenge we

encountered was for example the heterogeneity of the

terminology used to formulate FMs. Nonetheless, on

a more optimistic note, our findings also revealed

noteworthy similarities and common FMs that

support the need for a digital tool to aid departments

with their proactive risk assessment. Furthermore, the

level of safety awareness maturity strongly varies

among different RT organizations. In addition, the

knowledge about FMs is compartmentalized; if a

safety analysis was conducted somewhere, its results

usually stay in the department and are not widely

disseminated. Moreover, as technology in RT rapidly

advances, new, previously unidentified risks are

continuously emerging, presenting challenges to

professionals tasked with their implementation

(Ortiz-Lopez, 2009). Finally, education material and

non-proprietary digital tools to assist safety analysts

are practically non-existent.

The cumulative effect of all these challenges is that

incidents and errors of suboptimal RT treatment still

occur on daily basis (Ford and Evans, 2019). Hence,

there is an urgent need for assistance in conducting

hazard analyses, with the ultimate goal of enhancing

the safety of RT patients worldwide. This imperative

served as the driving force behind the initiation of the

i-SART project, a collaborative effort between the

computer science department at the Vrije Universiteit

in The Netherlands and the biomedical sciences

department at the University of West Attica in

Greece. The project addressed the following research

questions:

RQ1. Can we build an open-source software tool to

assist RT practitioners in conducting an effective

FMEA?

RQ2. Can we use machine learning to augment the

data obtained by the FMEA studies ?

3 i-SART, AN INTELLIGENT

FMEA ASSISTANT

The main goal of i-SART was to engage RT

practitioners in a dynamic FMEA learning

experience. Given the fact that RT professionals may

or may not have experience in FMEA, we expected

the usage of this tool to vary accordingly. On the one

HEALTHINF 2024 - 17th International Conference on Health Informatics

422

hand, FMEA beginners can use it as an expert system

to guide their analysis. On the other hand, RT experts

who are proficient in FMEA will be able to learn

about new FMs reported by other departments, or

share interesting FMs they have identified in their

institutions.

Technically, i-SART is designed as a cloud-

hosted web-application with two kinds of users:

administrators and RT practitioners (users), each with

their permissions and rights. The high-level

architecture of i-SART is illustrated in Fig. 3. Central

to i-SART is a novel database that aims to contain as

many as possible FMs that can occur in various RT

techniques, such as Intensity Modulated

Radiotherapy (IMRT), Stereotactic Body Radiation

Therapy (SBRT), etc. In the database, the FMs are

also grouped per sub-process and step in the RT sub-

process, such as Treatment planning, Treatment

delivery, etc. as illustrated in Fig. 1.

The web-application’s back-end was programmed

using Python 3.9 and the Django REST

framework.

Care was taken to ensure a secure transfer of

information between client and server, using the

JSON Web Token authentication

. A relational

database MySQL technology was used to achieve

persistent storage of both users’ and FM information.

The front-end user interface of i-SART was

developed using a Java-script-based framework

called Vue.js

. Its main function is to allow an RT

user to search in i-SART for FMs that might happen

in their particular RT process. There are also

searching, filtering and ordering functions available.

For example, if a user discovers a new FM in their

daily practice, they can add it to the i-SART database,

after a preliminary validation by the system

administrator, who is an RT expert. The administrator

can also visualize on a dashboard all the FMs in

different pie-charts, categorized based on their

severity or risk (see Fig. 4). A Vue chart component

library called Vue-ECharts

was used to plot these

charts. Finally, all users have the possibility to

evaluate the tool and send feedback and suggestions

to improve it.

https://www.djangoproject.com/

https://django-rest-framework-simplejwt.readthedocs.io/

en/latest/

4 FIRST RESULTS

4.1 First FM Data

The first step after the skeleton of the i-SART

prototype was ready, was to populate the FM

database. First 584 FMs were extracted from

scientific papers and un-published safety reports, all

written in English (see Table 1). All FMs were

classified into subprocesses and their corresponding

steps. Interesting is that we found that 32.5% of the

collected FMs fall into the subprocess Treatment

planning, 25.5% into the subprocess Treatment

delivery, and 20.7% into subprocess Imaging.

Table 1: Sources for the first collection of FMs .

Reference type #FMs Author/Year/Country

Research paper

16 Gilmore et al. (2021), UK

284 Pobbiati et al. (2019), Italy

220 Huq et al. (2016), USA

7 Bright et al. (2022), USA

26 Gehad et al. (2021), Egypt

Report from

individual RT

department

10 Not published, UK, 2022

21 Not-published, UK, 2021

Figure 3: The high-level architecture of i-SART.

https://vuejs.org/

https://github.com/ecomfe/vue-echarts

i-SART: An Intelligent Assistant for Safety Analysis in Radiation Therapy

423

Figure 4: A screenshot of the administrator’s dashboard

screen in i-SART.

4.2 Cleaning Data to Eliminate

Duplicates

Very soon however, we discovered that these “raw”

FMs contained many duplicates and ambiguities,

inherent to any text formulated in natural language.

We distinguished two types of duplicates: explicit,

defined as two exactly same FMs, which were easy to

detect, and implicit duplicates, where the semantics

was the same, but the syntax was different, which

were more difficult to detect. For example, we

considered the following two FMs as implicit

duplicates.

Collision risk due to gantry rotation

Gantry collision with visual aid device

We detected implicit duplicates using both manual

review by our RT expert team, and automated Natural

Language Processing (NLP) algorithms. For

example, first, the RT experts extracted a group of

keywords such as wrong, poor, imperfect, to help to

identify potential duplicates. Next, for each keyword,

all kinds of forms (i.e., verb, adjective, noun, adverb,

singular, plural) were generated using two NLP

libraries called inflect

and word-forms

and finally

were added to the keyword list. While we were aware

that words like "poor" and "wrong" may have

different causes and effects, we considered two FMs

containing these words as candidates for duplicates.

We eventually classified them as real duplicates only

after a thorough validation by our RT experts.

Next, we inspected FMs that exhibited a tree-like

structure. For instance, let us take a look at the

https://pypi.org/project/inflect/

following FMs, belonging to the subprocess Imaging

and its step “CT image acquisition”:

Wrong CT scan for treatment planning: wrong choice

of anatomical volumes

Wrong CT scan for treatment planning: fiducial

markers not implanted

Wrong CT scan for treatment planning (Vero®):

Optoelectronic markers not completely included into

the scan

Wrong CT scan for treatment planning

(CyberKnife®): scan volume not compliant to the

specifics requirements of the TPS

In all these four FMs, the text to the left of the colon

(:) describes the same unsafe situation (Wrong CT

scan), whereas the text to the right of the colon is an

elaboration on the specific causality. We suspected

that these FMs might be treated as implicit duplicates,

or at least be clustered in the same FM group. Again,

this happened in reality only after a validation by the

RT experts.

As a result of all these data-cleaning procedures,

we were able to eliminate 57 explicit duplicates, 37

implicit duplicates based on keywords and 130

implicit duplicates based on the tree-like structure.

Given the fact that an FM can be flagged as duplicate

multiple times by different filtering methods, we took

action to ensure each FM only appears once. As a

result, the total amount of uniquely duplicated FMs

was reduced to 166. Eventually, we ended up with

584-166 = 418 unique FMs, which we entered into the

database. We have to note that this is an indication of

the number of FMs available so far. This work is in

progress. Our team is working on fine-tuning the

application and improving the database. A prototype

of i-SART with approximately this number of FMs

will be soon available to be used and tested by

interested RT practitioners. To conclude, we would

like to emphasize that during the process of

eliminating the duplicates, the final decisions need to

be taken by our RT experts’ team, who will ensure

that no critical FM will get excluded by mistake.

4.3 Augmenting the Database with

Synthetic FMs

Although we were initially satisfied with the way we

populated the i-SART database, we also investigated

the possibility of augmenting the database with new,

synthetic AI-generated FMs. The reason for this was

the consideration that if the database will be used in

https://pypi.org/project/word-forms/

HEALTHINF 2024 - 17th International Conference on Health Informatics

424

the future for training machine learning algorithms, a

few hundred FMs will be definitely not sufficient to

achieve a high prediction accuracy. For example,

deep learning models used in NLP use typically

training datasets containing millions of items (Bailly,

2022). In this section, we will present a few

interesting results. We have to note here that these

were separate experiments and none of these results

have been yet implemented in i-SART.

Generative AI (GenAI) is a modern, powerful

technology that can produce new plausible media

content from existing content, including text, images,

audio, etc trying to mimic human creativity. It

originates in the research done at Google in 2017

(Vaswani, 2017) that first analyzed a language trying

to discover patterns in it, and then transformed this

analysis into a prediction on which word or phrase

should come next. Many GenAI algorithms exist,

varying from the probabilistic Naïve Bayes Networks

and Markov Models, to all kinds of feature-based

neural-networks variations, such as recursive neural

networks (RNN), convolutional neural networks

(CNN), and the GPT-2, -3 and -4 series, where GPT

stands for “Generative Pre-trained Transformer”.

Regardless the algorithm, automated text generation

works basically in the same way. In the beginning, all

probabilities or adjustable weights in the neural

network are unknown; we say that the model is not

trained. However, the model can learn these

parameters if provided with a huge number of training

examples. Eventually, when the training is finished,

and one starts with one word (also called prompt), the

model will be able to accurately predict the most

likely next word in a phrase.

Therefore, GenAI seemed a perfect approach

suitable for our purpose. We had a rather small

collection of training text data (the FMs) and we

wanted an AI algorithm to learn how to create more,

synthetic FMs. In line with these thoughts, we

conducted two preliminary experiments that explored

the performance of different GenAI algorithms.

The first experiment, in the context of a MSc CS

graduation project (Haddou, 2022), used two

different algorithms, Markov Chains and ChatGPT-3

to learn how to create new FMs based on an existing

collection. The training database was slightly

different, containing around 600 FMs collected from

literature and a few RT departments in Europe. From

these, eleven FMs that were generated with the

Markov Chains algorithm were presented for

validation to an RT expert. Out of these six were

found useful. There was at least one artificial FM

with a high RPN, namely “Incorrect image data set

associated with patient shifts determined” that was

interpreted by the RT expert as “patient shifts

determined by incorrect image data set”. Another FM

was very interesting because the RT expert had seen

it a lot of times before, namely ‘Patient head’s

position is not ideal’. The RT expert noted that this

FM would never come spontaneously to her mind.

This FM was clearly and correctly generated by the

Markov Chain algorithm.

The GPT-3 algorithm generated eleven FMs that

were also presented to the same RT expert. Out of these,

four of them were found useful. Especially the FM:

“Patient or nurse falls” and “Patient falls down due

to mobile phone dropping on the floor” were very

interesting. We could conclude from here that

synthetic FMs have the potential to raise awareness

or reveal unpredicted hazards that might occur during

any process, not necessarily RT specific.

The second experiment used a Generative

Adversarial Network (GAN) algorithm to generate

artificial FMs (Brophy, 2023). As a training dataset

we used our most recent FM database. GANs are a

branch of GenAI algorithms that consist of two

artificial neural networks, called Generator and

Discriminator (Goodfellow, 2020). The Generator

tries to generate new data as similar to original data

as possible, while the Discriminator’s role is to

determine if the input belongs to the real dataset or

not. The optimization process is characterized as a

game where the generator successfully learns to

“fool” the discriminator in such a way that the

discriminator cannot distinguish between the real one

and the synthetic one. In particular, our experiment

used the seqGAN model (Yu, 2017).

The Bilingual Evaluation Understudy (BLEU)

score was one of the metrics used to measure the

quality of the FMs produced by the generative

algorithm. The basic idea of the BLEU score is

straightforward: the closer the synthetic FM is to the

human-generated target sentence, the better it is; a

score of 1 means a perfect match, and 0 means no

match. A BLEU-score has different levels (n),

depending on how many n-grams are being

compared. For example, if n=1, each word from the

original and synthetic text will be compared, and if

n=2, each word pair will be matched. As training

data we used all the 584 raw FMs initially collected

as described in section 4.1, plus 1721 FM taken from

the headlines of incidents reported in IAEA

SAFRON, (SAFety in Radiation ONcology), an

i-SART: An Intelligent Assistant for Safety Analysis in Radiation Therapy

425

international platform that collects RT incidents and

promotes patient safety

The seqGAN model implemented in Pytorch

was

able to create additional 640 artificial FMs. From

these, 230 were identified as useful by an RT expert,

with 53 duplicates. A handful of them (only 9) were

considered as really novel with respect to the existing

FMs in the database (see Table 2). However, so far

we found that the synthetic FMs lacked syntactic

accuracy and clarity. Fig. 5 shows the performance of

the seqGAN algorithm measured using the BLEU

scores with levels n = 2, 3, 4, 5. The black dot line

splits the training into two phases: 1) Before the

divider is the pre-training process where the initial

generator was trained, and 2) after the divider is the

adversarial training process where the generator

continues to update based on the reward from the

discriminator. We can see that the more grams the

calculation of a BLEU score is based on, the lower

the score. In our experiment, the BLEU-2 values were

the highest and reached the maximum

value of 0.6.

Figure 5: The BLEU-[2, 3, 4, 5] scores of the synthetic FM

generation using the seqGAN model.

These first results show that using GenAI algorithms

is an interesting idea to generate synthetic FM.

However, more efforts will be needed in future to

increase their accuracy and eliminate the need of

human judgement.

Finally, we also identified a few limitations to this

approach. For example, when deciding if the newly

generated FM is a valid one, we consulted only one

RT practitioner, while the assessment of any FM in a

process needs an RT team. Moreover, we didn’t

include the steps in which an FM could occur. This

would bring more clarity to the results. We also

expect that a larger FM dataset will also improve the

accuracy. This will happen in time, when i-SART will

be used by a large community of RT practitioners.

Moreover, while AI integration is innovative, there’ s

https://www.iaea.org/resources/rpop/resources/databases-

and-learning-systems/safron

a risk of overconfidence in AI-generated FMs without

adequate human oversight.

Table 2: Novel, synthetic FMs created using GANs.

Synthetic Failure mode Correction/Comments

RT ex

ert

incorrect isocentre not

use

incorrect isocentre used

patient positioning with

wrong tattoo

(Nice! I haven’t seen this

in the 584 FMs!)

patient was treated with

wron

side of tattoo

(sounds similar to

revious one

)

incorrect selection of

appointments delivery

recorded on verification

stem

incorrect selection of

appointments on record

and verify system

PTV received higher and

treated on the patient

A higher dose was

prescribed for the PTV*

and was delivered to the

patient

incorrect collimator

angles not imported

collimator angles not

imported (Yes, although

I’m not sure if this is

technicall

ossible.

)

old treatment protocol use of the old treatment

protocol instead of the

new one

wrong field size on portal

ima

this is fully correct

planned for the wrong

beam angles for one of

treatment fields

wrong beam angles for

one of treatment fields

*PTV means Planning Target Volume and is the region

around the tumor that needs to be irradiated

5 CONCLUSION AND FUTURE

WORK

We presented i-SART, a novel web-application that

aims to assist RT practitioners in conducting a sound

proactive safety assessment using FMEA. i-SART is

the result of a successful cooperation between RT and

computer science researchers. Central is an FM

database that can grow due to contributions from

participating users. We also experimented with

machine learning techniques, such as NLP for

duplicates elimination and generative AI to create

synthetic FMs. We conclude that although machine

learning can be useful in assisting a safety assessment

process, the results need to be always validated by RT

experts. Future work includes optimizing the machine

learning algorithms, including a variety of other

https://github.com/williamSYSU/TextGAN-PyTorch

HEALTHINF 2024 - 17th International Conference on Health Informatics

426

safety analysis methods besides FMEA and

investigating the possibilities to offer i-SART as an

open-source collaborative tool for the international

RT community with the common goal of contributing

to a safe and fair global healthcare.

ACKOWLEDGEMENTS

We would like to thank prof. Annette ten Teije for her

support in pursuing our research. A prototype of i-

SART is currently hosted in the SciCloud, a virtual

server service for research offered by the Vrije

Universiteit in Amsterdam.

REFERENCES

Arnzen, H.E. (1966). Failure Modes and Effect Analysis:

Powerful Engineering Tool for Component and System

Optimization. Reliability and Maintainability

Conference, 355-71.

Bailly, A. et al. (2022). Effects of dataset size and

interactions on the prediction performance of logistic

regression and deep learning models. Computer

Methods and Programs in Biomedicine, Volume 213,

106504.

Bright, M. et al. (2022). Failure modes and effects analysis

for surface-guided DIBH breast radiotherapy. Journal

of applied clinical medical physics, 23(4), e13541

Borrás, C. et al. (2006). Overexposure of radiation therapy

patients in Panama: problem recognition and follow-up

measures. Revista Panamericana de Salud Pública,

20(2-3), 173-187.

Brophy, E. et al. (2023). Generative Adversarial Networks

in Time Series: A Systematic Literature Review. ACM

Comput. Surv. 55, 10, Article 199.

Council of the European Union. (2014). Directive

2013/59/Euratom on basic safety standards for

protection against the dangers arising from exposure to

ionising radiation and repealing Directives

89/618/Euratom, 90/641/Euratom, 96/29/Euratom,

97/43/Euratom and 2003/122/Euratom, Official

Journal of EU. L13; 57:1-73.

deRosier, J. et al. (2002). Using health care failure mode

and effect analysis. Joint Commission J. Qual.

Improvement 27, 248–267.

Ford, E. and Evans, S.B. (2018). Incident learning in

radiation oncology: A review, Medical Physics, 45:

e100-e19.

Gehad, S.M. et al. (2021). Failure Mode and Effect

Analysis for Three Dimensional Radiotherapy at Ain

Shams University Hospital, Int. J. of Adv. Res. 9 (Dec).

971-983.

Gilmore, M. D. and Rowbottom, C. G. (2021). Evaluation

of failure modes and effect analysis for routine risk

assessment of lung radiotherapy at a UK center. Journal

of applied clinical medical physics, 22(5), 36-47.

Goodfellow, I. et al. (2020). Generative adversarial

networks. Communications of the ACM, 63(11), 139-

144.

Haddou, S. 2022. iSART: Digital safety assistant for the

radiotherapy process. Msc CS thesis, Vrije Universiteit

Amsterdam.

Huq, M. S. et al. (2016). The report of Task Group 100 of

the AAPM: Application of risk analysis methods to

radiation therapy quality management, Medical

Physics, 43: 4209-62.

Ibanez-Rosello, B. et al. (2018). Failure modes and effects

analysis of total skin electron irradiation technique.

Clinical and Translational Oncology, 20(3), 330-365.

International Atomic Energy Agency (IAEA). (2018).

Radiation protection and safety in medical uses of

ionizing radiation. Safety Standards Series No. SSG-

46;

Kane, P., The impact of the VERT virtual reality system on

teaching and learning associated with radiation therapy

planning skills in the second year of the Bachelor of

Radiation Therapy. 2014

Leveson, N.G. and Turner, C.S. (1993). An investigation of

the Therac-25 accidents, in Computer, vol. 26, no. 7,

18-41.

National Cancer Institute (NCI). (2017). Cancer Statistics,

https://www.cancer.gov/, Accessed online 7-10-2023.

Olch, A. (2014). Quality and Safety in Radiation therapy

Learning the New Approaches in Task Group 100 and

Beyond, Medical Physics, 41: 067301.

Ortiz-Lopez, P. et al. (2009). Preventing accidental

exposures from new external beam radiation therapy

technologies. Ann ICRP, 39(4):1e86.

Pawlicki, T. et al. (2011). Quality and Safety in Radiation

therapy (CRC Press).

Pawlicki, T. et al. (2016). Application of systems and

control theory-based hazard analysis to radiation

oncology. Medical Physics, 43(3):1514-30.

Pobbiati, C. et al. (2019). Clinical risk analysis of the

patient’s path in an Advanced Radiotherapy Center

through FMEA method. Journal of Biomedical

Practitioners, 3(1).

Sarchosoglou, A. et al. (2022). Failure modes in stereotactic

radiosurgery. A narrative review, Radiography,

Volume 28, Issue 4, Pages 999-1009.

Silvis-Cividjian, N. et al. (2020). Using a systems-theoretic

approach to analyze safety in radiation therapy-first

steps and lessons learned. Safety Science, 122, 1-10.

Vaswani, A. et al. (2017). Attention is all you need. In

Proceedings of the 31st International Conference on

Neural Information Processing Systems (NIPS'17).

6000–6010.

Yu, L. et al. (2017). SeqGAN: Sequence generative

adversarial nets with policy gradient. In Proceedings of

the AAAI conference on artificial intelligence (31,1).

i-SART: An Intelligent Assistant for Safety Analysis in Radiation Therapy

427