Preparing Ultrasound Imaging Data for Artiﬁcial Intelligence Tasks:

Anonymisation, Cropping, and Tagging

Dimitrios Pechlivanis

1 a

, Stylianos Didaskalou

2 b

, Eleni Kaldoudi

1,2 c

and George Drosatos

2 d

School of Medicine, Democritus University of Thrace, 68100 Alexandroupoli, Greece

Institute for Language and Speech Processing, Athena Research Center, 67100 Xanthi, Greece

Keywords:

Ultrasound Imaging, DICOM, Anonymisation, Cropping, Tagging, Artiﬁcial Intelligence (AI).

Abstract:

Ultrasound imaging is a widely used diagnostic method in various clinical contexts, requiring efﬁcient and

accurate data preparation workﬂows for artiﬁcial intelligence (AI) tasks. Preparing ultrasound data presents

challenges such as ensuring data privacy, extracting diagnostically relevant regions, and associating contextual

metadata. This paper introduces a standalone application designed to streamline the preparation of ultrasound

DICOM ﬁles for AI applications across different medical use cases. The application facilitates three key pro-

cesses: (1) anonymisation, ensuring compliance with privacy standards by removing sensitive metadata; (2)

cropping, isolating relevant regions in images or video frames to enhance the utility for AI analysis; and (3)

tagging, enriching ﬁles with additional metadata such as anatomical position and imaging purpose. Built with

an intuitive interface and robust backend, the application optimises DICOM ﬁle processing for efﬁcient inte-

gration into AI workﬂows. The effectiveness of the tool is evaluated using a dataset of Deep Vein Thrombosis

(DVT) ultrasound images, demonstrating signiﬁcant improvements in data preparation efﬁciency. This work

establishes a generalizable framework for ultrasound imaging data preparation while offering speciﬁc insights

into DVT-focused AI workﬂows. Future work will focus on further automation and expanding support to ad-

ditional imaging modalities as well as evaluating the tool in a clinical setting.

1 INTRODUCTION

Machine learning (ML) has already achieved signif-

icant breakthroughs in medicine, especially in the

automated diagnosis in medical imaging, automated

management of administrative tasks, and exploration

of the structure of bio-molecules, particularly proteins

(Habehh and Gohel, 2021). A critical prerequisite

for the successful application of machine learning in

medical imaging is the preparation of high-quality, la-

belled datasets. This process, often referred to as ML

training data preparation, is particularly intricate due

to several factors: the need for strict anonymisation to

ensure patient privacy, the identiﬁcation and extrac-

tion of diagnostically relevant image regions, and the

association of contextual metadata, such as anatom-

ical position and purpose. Ensuring patient privacy

through anonymisation is especially critical, as it sup-

ports both ethical considerations and compliance with

data protection regulations in medical research.

https://orcid.org/0009-0001-3583-772X

https://orcid.org/0000-0001-5932-9626

https://orcid.org/0000-0002-6054-4961

https://orcid.org/0000-0002-8130-5775

Medical imaging and signal data management is

almost exclusively based on the DICOM (Digital

Imaging and Communication On Medicine) image

and communication standard (ISO 12052, 2017). DI-

COM supports all medical imaging modalities but

also organises data into two key components: (1)

the metadata, which include patient-speciﬁc informa-

tion (e.g., name, date of birth, and hospital details),

as well as examination-speciﬁc data (e.g., modality

type, acquisition parameters), and (2) the imaging

data, which consist of pixel-based representations of

the medical image or video. In addition, some meta-

data ﬁelds may be “burnt-in” directly onto the imag-

ing data, as is common in ultrasound imaging, further

complicating the anonymisation process.

Traditional approaches to DICOM data prepara-

tion for ML training, including the removal or ob-

fuscation of sensitive metadata and annotations from

both the header and the image itself, are often man-

ual and time-intensive. These limitations hinder the

scalability of AI-based solutions in clinical practice,

underlining the importance of automated tools that

can handle both anonymisation and efﬁcient dataset

preparation. This involves ensuring data privacy, stan-

dardising formats, isolating relevant image regions,

Pechlivanis, D., Didaskalou, S., Kaldoudi, E. and Drosatos, G.

Preparing Ultrasound Imaging Data for Artiﬁcial Intelligence Tasks: Anonymisation, Cropping, and Tagging.

DOI: 10.5220/0013379400003911

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

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 2: HEALTHINF, pages 951-958

ISBN: 978-989-758-731-3; ISSN: 2184-4305

951

and enriching the data with metadata to meet the qual-

ity requirements for ML training.

Ultrasound imaging, as a widely used modality in

clinical practice, exempliﬁes these challenges due to

its operator-dependent variability and its application

across diverse diagnostic contexts. While these issues

are common across ultrasound imaging use cases,

they become particularly critical in speciﬁc condi-

tions such as deep vein thrombosis (DVT). DVT is

a serious medical condition characterised by the for-

mation of blood clots within deep veins, most com-

monly in the lower extremities (Waheed et al., 2024).

Ultrasound imaging is the primary diagnostic modal-

ity for DVT due to its non-invasive nature and real-

time imaging capabilities. However, the complex-

ity of analysing ultrasound data, coupled with vari-

ability in operator technique and patient anatomy,

poses challenges for consistent and accurate diagno-

sis (Baker et al., 2024). Artiﬁcial intelligence (AI)

offers a promising approach to addressing these chal-

lenges, automating tasks such as vessel segmenta-

tion and thrombus detection (Maiti and Arunachalam,

2022).

This paper introduces a tool, called

US-DICOMizer, to automate and streamline the

preparation of ultrasound DICOM ﬁles for AI-based

workﬂows. While the tool is designed for general

ultrasound imaging applications, it is evaluated using

a dataset of DVT ultrasound exams as a test case.

The application incorporates three key functionali-

ties: (1) anonymisation to remove sensitive patient

information while preserving essential metadata for

AI tasks, (2) cropping to extract relevant regions

from images or videos, and (3) tagging to annotate

ﬁles with critical metadata, such as anatomical

position, imaging purpose, and other contextual

information. These functionalities aim to address

the unique requirements of ultrasound imaging data

preparation while ensuring compliance with data

privacy regulations and clinical standards.

This work is part of the ThrombUS+ project, co-

funded by the European Union (EU) under Grant

Agreement No. 101137227 (Kaldoudi et al., 2024).

The ThrombUS+ initiative aims to develop an innova-

tive, operator-independent wearable diagnostic device

designed to facilitate the early detection and diagno-

sis of DVT using various diagnostic data modalities,

including ultrasound imaging. As part of this initia-

tive, the US-DICOMizer application addresses criti-

cal data preparation needs for AI-based workﬂows in

DVT and other ultrasound imaging applications.

The remainder of this paper is organised as fol-

lows: Section 2 provides an overview of related work

in DICOM data preparation and AI in ultrasound

imaging. Section 3 details the methodology and

design of the application, including its anonymisa-

tion, cropping, and tagging functionalities. Section 4

discusses the technical implementation details, high-

lighting the tools and frameworks employed. Section

5 presents results demonstrating the performance and

the user experience of the application. Finally, Sec-

tion 6 concludes the paper by summarising the contri-

butions of this work to AI-based medical imaging and

provides future directions.

2 RELATED WORK

The preparation of ultrasound imaging data for artiﬁ-

cial intelligence (AI) tasks requires specialised tools

to address the unique challenges of this modality.

Anonymisation, cropping, and tagging are critical

steps to ensure patient privacy, enhance data utility,

and facilitate effective data sharing. Existing solu-

tions have focused on general medical imaging or

speciﬁc modalities, often neglecting the particular re-

quirements of ultrasound imaging for tasks like vessel

segmentation and thrombus detection. This section

reviews related work on anonymisation, cropping,

and tagging techniques in medical imaging, highlight-

ing their relevance and limitations in the context of

ultrasound.

Numerous anonymisation tools have been devel-

oped to safeguard patient privacy while ensuring com-

pliance with DICOM (Digital Imaging and Commu-

nications in Medicine) standards. Monteiro et al.

(2017) introduced a de-identiﬁcation pipeline lever-

aging convolutional neural networks (CNNs) for char-

acter recognition, achieving an anonymisation rate of

89.2%. Similarly, Rodr

ıguez Gonz

alez et al. (2010)

developed an open-source toolkit for DICOM data

de-identiﬁcation tailored to multicenter trials, offer-

ing customisation based on speciﬁc privacy needs.

However, these tools are primarily designed for neu-

roimaging and lacked features speciﬁc to ultrasound

workﬂows. Haselgrove et al. (2014b) presented a sys-

tem for anonymising DICOM data in neuroimaging

research, ensuring compliance with quality control

standards but leaving the unique needs of ultrasound

imaging unaddressed.

Cropping and tagging techniques play a cru-

cial role in isolating diagnostically relevant regions,

anonymising burned-in annotations text, and asso-

ciating metadata with medical images. Interactive

anonymisation tools, such as the Cornell Anonymi-

sation Toolkit (Xiao et al., 2009), allow users to man-

ually crop and anonymise datasets but do not offer

automation or optimisation for ultrasound imaging.

Tools like DICOM Anonymiser and dicom-anon4

support tagging within the framework of DICOM

standards (Haselgrove et al., 2014a), but their focus

on neuroimaging limits their relevance to ultrasound-

speciﬁc needs. More recently, image cropping

techniques based on grid-anchor formulations have

HEALTHINF 2025 - 18th International Conference on Health Informatics

952

emerged in general computer vision (e.g., CNN-based

methods), but these approaches are designed for aes-

thetic image processing rather than clinical applica-

tions (Zeng et al., 2022).

Despite these advancements, signiﬁcant gaps re-

main in addressing the speciﬁc requirements of ul-

trasound imaging. Existing tools are often modality-

speciﬁc and fail to provide comprehensive solutions

that integrate anonymisation, cropping, and tagging

for ultrasound data. Furthermore, inconsistent stan-

dards across imaging communities pose challenges to

the adoption of uniﬁed workﬂows. This paper ad-

dresses these challenges by introducing a tool, as a

standalone application, tailored to the preparation of

ultrasound DICOM data for AI tasks. While the tool

is designed to be broadly applicable across ultrasound

imaging use cases, its functionality is evaluated using

deep vein thrombosis (DVT) examinations as a test

case to validate its performance and utility.

3 METHODOLOGY

The development of US-DICOMizer is driven by

the broader requirements of ultrasound imaging data

preparation for artiﬁcial intelligence (AI) tasks, while

incorporating insights from the ThrombUS+ project

(Kaldoudi et al., 2024). As part of this initiative,

Clinical Study A (Drougka, 2024), described in de-

tail in the data collection manual available at https:

//app.thrombus.eu/studies/a/, is used to validate its

functionality. The methodology combines insights

from established imaging protocols with tailored tool

development to streamline key data preparation steps,

such as anonymisation, cropping, and tagging. It ex-

cludes the labelling and annotation of the collected

images and videos, which will be performed as a sep-

arate and necessary step for AI training.

Requirements and Design Principles. Clinical

Study A outlines a structured process for ultrasound

data collection, emphasising accurate imaging of spe-

ciﬁc anatomical sites, tagging critical metadata, and

ensuring privacy compliance. This process inspired

the design of US-DICOMizer, following a user-

centred approach which automates data preparation

to reduce manual effort while preserving the integrity

and diagnostic value of the images and videos. The

tool adheres to the following principles:

1. Data Privacy Compliance: Adherence to

anonymisation standards, ensuring compatibility

with HIPAA and GDPR.

2. Standardisation and Consistency: Integration

of tagging functionalities aligned with DICOM

standards to support AI-based downstream tasks.

3. Workﬂow Efﬁciency: Minimising manual inter-

vention while enabling ﬂexibility through interac-

tive features.

4. Automation and Scalability: Streamlining

repetitive tasks such as cropping and anonymisa-

tion to improve efﬁciency across large datasets.

5. Clinical Relevance: Supporting tagging based on

the anatomical positions and diagnostic context

detailed in the Clinical Study A protocol.

Data Collection Context. In Clinical Study A,

healthcare professionals collect ultrasound images

and videos following a structured protocol to assess

key venous sites in the lower extremities. The imag-

ing included:

• Compression Ultrasound: Capturing frames and

videos to evaluate vein compressibility at speciﬁc

sites (e.g., common femoral vein, popliteal vein).

• Color Doppler Ultrasound: Enhancing diagnostic

detail for speciﬁc anatomical regions.

• Optional Imaging: Capturing pathological sites

(e.g., thrombus visualisation) where necessary.

US-DICOMizer integrates these requirements by

providing a robust framework for importing, crop-

ping, tagging, anonymizing, and exporting DICOM

ﬁles, preserving the diagnostic and research value of

the data.

Workﬂow. The workﬂow followed for preparing DI-

COM ﬁles with US-DICOMizer is presented in Fig-

ure 1 and involves the following key steps:

1. File Loading and Validation: Users can load sin-

gle DICOM ﬁles, entire folders, or compressed

archives. The tool validates the ﬁles to ensure

compatibility and displays their metadata for re-

view. Multiframe DICOMs are previewed frame-

by-frame to facilitate inspection.

2. Cropping and Tagging:

• Cropping: Users deﬁne relevant regions of

interest either manually through an interac-

tive GUI or by applying preset conﬁgurations.

Multiframe cropping ensures uniformity across

videos.

• Tagging: Metadata, including anatomical po-

sition (e.g., proximal, distal), the side (left or

right), and the image purpose, are applied to

each ﬁle.

3. Anonymisation: Sensitive information is auto-

matically removed or replaced, with user-deﬁned

conﬁgurations specifying which tags are to be

deleted. The tool generates unique identiﬁers for

studies, series, and instances to ensure compliance

with data privacy regulations.

4. Export: Prepared DICOM ﬁles are exported as a

ZIP archive, maintaining a standardised structure

and ensuring compatibility with AI workﬂows.

Preparing Ultrasound Imaging Data for Artiﬁcial Intelligence Tasks: Anonymisation, Cropping, and Tagging

953

List of

available

DICOM files

Load and

validate DICOM

files

Cropping

parameters

Selection of

required

DICOM files

Final list of

DICOM files

Tags

Apply and

adjust cropping

region

Select tag

Apply

anonymisation

ZIP of

anonymized

DICOM files

Attributes to

be deleted

Export

anonymised

DICOM files

Figure 1: US-DICOMizer workﬂow.

By aligning the application workﬂow with this

protocol, US-DICOMizer bridges the gap between

clinical imaging practices and AI-driven research, en-

suring standardised, high-quality datasets.

4 IMPLEMENTATION DETAILS

The development of the US-DICOMizer focuses on

facilitating the preparation of ultrasound DICOM ﬁles

for artiﬁcial intelligence (AI) tasks through an in-

teractive and semi-automated workﬂow. The ap-

plication is open-source and available on GitHub

at https://github.com/thrombusplus/US-DICOMizer.

Below, we outline the technical details of its design

and implementation.

Development Environment and Frameworks. The

application is implemented in Python, leveraging the

following key libraries:

• Tkinter: For building a graphical user interface

(GUI) to facilitate intuitive user interactions.

• Pydicom: For reading, editing, and anonymising

DICOM ﬁles.

• SimpleITK: For image processing and visualisa-

tion of DICOM images.

• Pillow (PIL): For handling image formats and

performing operations like cropping and saving.

• NumPy: For efﬁcient array manipulations of im-

age data.

• Matplotlib: For visualising images during pro-

cessing.

• Scikit-Image: For advanced image manipulation

and processing tasks.

Core Functionalities. Users can load individual DI-

COM ﬁles, entire folders, or ZIP archives containing

multiple DICOM ﬁles. Imported ﬁles are displayed in

a tree-view structure within the application, allow-

ing users to preview metadata and image content. For

multiframe DICOMs, a video slider facilitates frame-

by-frame visualisation.

Cropping regions are deﬁned interactively through

the GUI or automatically based on pre-conﬁgured set-

tings in the settings.ini ﬁle. These cropping pa-

rameters can be applied uniformly to all frames of

multiframe DICOMs. Metadata tagging ensures the

association of critical information, including anatom-

ical position and image purpose, with each ﬁle. Crop

settings are also saved to the conﬁguration ﬁle for

consistency across sessions.

Sensitive patient identiﬁers are removed or re-

placed during the anonymisation process. At-

tributes speciﬁed in the conﬁguration ﬁle under the

[attributes to del] section are deleted from each

DICOM ﬁle. The application also generates new

unique identiﬁers (UIDs) for the study, series, and in-

stances to maintain data integrity while ensuring pri-

vacy compliance.

Processed DICOM ﬁles are exported as a com-

pressed ZIP archive. The naming conventions and

folder structure of the output are customisable via

the application settings, ensuring compatibility with

downstream workﬂows.

Error Handling and Logging. The application in-

corporates robust error handling to prevent workﬂow

interruptions. For instance:

• Validation of DICOM ﬁles ensures that non-

DICOM inputs are skipped during batch process-

ing.

• Invalid or missing cropping parameters result in

detailed error messages, guiding users to update

their conﬁguration ﬁles.

• A dedicated logging system records application

activities and errors in a structured format for de-

bugging and audit purposes.

Customisability. User-deﬁned settings, such as crop-

ping areas, compression levels, and attributes removal

lists, are stored in an INI conﬁguration ﬁle. This ﬁle

allows for ﬁne-grained control over the application’s

behaviour, enabling customisation to suit various use

cases and device conﬁgurations.

HEALTHINF 2025 - 18th International Conference on Health Informatics

954

(a) Loading of DICOM ﬁles. (b) Selecting the appropriate DICOM ﬁles for anonymisation.

Figure 2: Loading, selecting and previewing DICOM ﬁles.

Performance Optimisation. The application em-

ploys multithreading for tasks like previewing images

and processing large datasets, ensuring smooth opera-

tion without GUI freezes. Efﬁcient handling of image

data through NumPy arrays and PIL ensures minimal

processing time.

5 RESULTS

The proposed US-DICOMizer application is evalu-

ated for its ability to streamline the preparation of ul-

trasound DICOM ﬁles for artiﬁcial intelligence (AI)

tasks through an intuitive sequence of steps: ﬁle load-

ing and preview, cropping and tagging, anonymisa-

tion, and export. Below, we describe the workﬂow

and outcomes achieved during testing.

Workﬂow Evaluation. Initially, users start the pro-

cess by importing DICOM ﬁles through multiple op-

tions, including selecting individual ﬁles, importing

entire folders, or uploading compressed archives (Fig-

ure 2.a). Imported ﬁles are displayed in the “Se-

lected File” list, allowing users to preview images and

inspect DICOM attributes, such as dimensions and

metadata (Figure 2.c). Multiframe DICOM ﬁles in-

clude an interactive video slider for frame-by-frame

Preparing Ultrasound Imaging Data for Artiﬁcial Intelligence Tasks: Anonymisation, Cropping, and Tagging

955

inspection. The interface provides users with the ﬂex-

ibility to reorder, remove, or review ﬁles before fur-

ther processing (Figure 2.b).

Once ﬁles are selected, users deﬁne cropping re-

gions to isolate diagnostically relevant portions of the

image or video frames. Predeﬁned cropping settings

are automatically applied where applicable, with the

option for manual adjustment using an interactive pre-

view tool (Figure 3.a). Cropping parameters could be

applied uniformly across all frames or adjusted for

individual ﬁles, ensuring consistency and relevance

in data preparation. Additionally, cropping parame-

ters are saved to a conﬁguration ﬁle, allowing users

to replicate consistent settings across multiple ses-

sions (Figure 3.b). Concurrently, users add tags to

each ﬁle to specify anatomical position (e.g., femoral

vein (FV) or popliteal vein (PV)), side (e.g., left or

right leg), and image purpose (e.g., diagnostic or non-

diagnostic quality) (Figure 3.c). These tags ensure

the association of critical metadata with the prepared

ﬁles, enhancing their utility for downstream AI tasks.

(a) Process of image cropping.

(b) Save cropping area in settings. (c) Select tag.

Figure 3: Cropping and tagging process.

With cropping and tagging completed, the

anonymisation process is initiated (Figure 4). The

application removes sensitive patient identiﬁers while

preserving essential metadata for AI processing. Files

with correctly deﬁned cropping areas and tags are vi-

sually marked to indicate readiness for anonymisa-

tion. Validation checks prevented the anonymisation

of incomplete or improperly prepared ﬁles, maintain-

ing data integrity.

Anonymised DICOM ﬁles are exported as a com-

pressed ZIP archive for convenient sharing and stor-

age (Figure 5). Users specify the destination folder

for the exported archive, completing the preparation

process. This ﬁnal step ensures that the ﬁles are ready

for integration into AI workﬂows with minimal man-

ual intervention.

Performance and User Experience. The applica-

tion demonstrates high efﬁciency and ease of use.

On average, ultrasound exams comprising 15-30 DI-

COM ﬁles (including single-frame and multiframe)

are fully prepared (including loading, cropping, tag-

ging, anonymisation, and export) within 3-6 minutes.

Exporting to a ZIP archive required less than 1–2 sec-

onds for exams of similar size.

The anonymisation process was evaluated for

its average execution times per frame across var-

ious photometric interpretations and cropping ar-

eas, as shown in Table 1. For single-frame ultra-

sound images with a resolution of 1200x800 pixels,

Monochrome2 photometric interpretation required

19–23 milliseconds, depending on the cropping area.

In contrast, the YBR FULL 422 and RGB interpre-

tations took longer, with times ranging from 25–30

milliseconds and 49–61 milliseconds, respectively.

For ultrasound multiframe DICOMs, the processing

Table 1: Average execution times (per frame) of anonymi-

sation process for various photometric interpretations and

cropping areas.

Media

Type

(resolution)

Photometric

Interpretation

Cropping

Area

(pixels)

AVG Time

per Frame

(msec)

Ultrasound

Image

(1200x800)

[100 runs]

Monochrome2

576x432 19

768×576 21

1024×768 23

YBR FULL 422

576x432 25

768×576 27

1024×768 30

RGB

576x432 49

768×576 53

1024×768 61

Ultrasound

Multi-frame

(1200x800)

[10 runs]

Monochrome2

576x432 102

768×576 105

1024×768 108

YBR FULL 422

576x432 117

768×576 119

1024×768 124

RGB

576x432 241

768×576 244

1024×768 250

HEALTHINF 2025 - 18th International Conference on Health Informatics

956

Figure 4: Start the anonymisation process after properly cropping and tagging the required DICOM ﬁles.

Figure 5: Export anonymised DICOM ﬁles as a ZIP ﬁle.

times increased as expected: Monochrome2 required

102–108 milliseconds per frame, YBR FULL 422 re-

quired 117–124 milliseconds, and RGB interpretation

ranged from 241–250 milliseconds per frame.

These results were produced using ultrasound ex-

ams from eight different ultrasound medical devices

manufactured by four vendors, and calculated on a

laptop computer running Windows 10, with an In-

tel Core i7-7700HQ CPU at 2.8 GHz and 16 GB of

RAM. While DVT datasets were used for testing, the

tool is designed for broader ultrasound imaging use

cases. Conclusively, this experiment highlights that

performance trade-offs are associated with both pho-

tometric complexity and cropping dimensions.

User feedback emphasises the application’s intu-

itive interface and streamlined workﬂow. Key fea-

tures, such as interactive previews before and after

anonymisation as well as cropping and tagging tools,

contribute to user satisfaction. Keyboard shortcuts

for common tasks further enhanced efﬁciency, mak-

ing the application a practical tool for medical imag-

ing professionals.

6 CONCLUSIONS AND FUTURE

WORK

The development and evaluation of US-DICOMizer

underscore its effectiveness in streamlining the prepa-

ration of ultrasound DICOM ﬁles for artiﬁcial intelli-

gence (AI) tasks. By automating key processes such

as anonymisation, cropping, and tagging, the appli-

cation addresses critical challenges in ultrasound data

preparation, including compliance with privacy stan-

dards, relevance of imaging regions, and metadata

contextualisation. The integration of these function-

alities into an intuitive, open-source platform facili-

tates scalable, standardised workﬂows for various ul-

trasound imaging AI applications. While the tool was

validated using deep vein thrombosis (DVT) datasets,

its design and functionality are adaptable to a wide

range of clinical use cases.

Preliminary results demonstrate that US-

DICOMizer signiﬁcantly reduces manual effort

and enhances the consistency of prepared datasets,

enabling faster and more reliable AI model training.

The alignment of the tool’s design with the structured

protocol of Clinical Study A from the ThrombUS+

project ensures its clinical relevance and applicabil-

ity. Furthermore, its adaptability to various imaging

Preparing Ultrasound Imaging Data for Artiﬁcial Intelligence Tasks: Anonymisation, Cropping, and Tagging

957

workﬂows and compliance with industry standards

position it as a versatile tool for medical imaging

research.

Despite these achievements, several opportunities

for future development remain. First, expanding the

tool’s support to additional imaging modalities, such

as computed tomography (CT) or magnetic resonance

imaging (MRI), could broaden its utility. Second,

incorporating advanced AI algorithms directly into

the tool could enable automated cropping and tag-

ging based on learned patterns, further reducing the

need for manual input. Third, large-scale validation

of the tool in clinical settings is necessary to evaluate

its robustness and user satisfaction across diverse en-

vironments and datasets. Lastly, integrating real-time

feedback mechanisms and interoperability with eCRF

(electronic case report form) platforms could stream-

line data sharing, automate the collection process and

ensure compliance with clinical trial regulations.

In conclusion, US-DICOMizer represents a sig-

niﬁcant step toward automating and standardising

ultrasound data preparation for AI-driven health-

care solutions. With planned enhancements and

broader adoption, it has the potential to accelerate ad-

vancements in diagnostic imaging and personalised

medicine, supporting a wide range of clinical and re-

search applications.

ACKNOWLEDGEMENTS

This work is co-funded by the European Union, un-

der the Horizon Europe Innovation Action Throm-

bUS+ (Grant Agreement No. 101137227). Views

and opinions expressed are however those of the au-

thors only and do not necessarily reﬂect those of

the European Union or HADEA as the granting au-

thority. Neither the European Union nor the grant-

ing authority HADEA can be held responsible for

them. Also, this work was carried out in the context

of the Inter-Institutional Master’s Program “Biomed-

ical Informatics” with the support of the School of

Medicine, Democritus University of Thrace and the

Athena Research Center in Greece.

REFERENCES

Baker, M., Anjum, F., and dela Cruz, J. (2024). Deep ve-

nous thrombosis ultrasound evaluation. In StatPearls.

StatPearls Publishing, Treasure Island (FL).

Drougka, I. (2024). D7.1: Clinical studies initiation pack-

age – Study A. ThrombUS+ Project, https://doi.org/

10.5281/zenodo.11371924.

Habehh, H. and Gohel, S. (2021). Machine learning in

healthcare. Current Genomics, 22(4):291–300.

Haselgrove, C., Poline, J.-B., and Kennedy, D. N. (2014a).

Comment on “A simple tool for neuroimaging data

sharing”. Frontiers in Neuroinformatics, 8.

Haselgrove, C., Poline, J.-B., and Kennedy, D. N. (2014b).

A simple tool for neuroimaging data sharing. Fron-

tiers in Neuroinformatics, 8.

ISO 12052 (2017). Health informatics – Digital imaging

and communication in medicine (DICOM) including

workﬂow and data management. Standard, Interna-

tional Organization for Standardization.

Kaldoudi, E., Marozas, V., Rytis, J., Pousset, N., Legros,

M., Kircher, M., Novikov, D., Sakalauskas, A., Mous-

takidis, P., Ayinde, B., Moltani, L. A., Balling, S.,

Vehkaoja, A., Oksala, N., Macas, A., Balciuniene,

N., Bigaki, M., Potoupnis, M., Papadopoulou, S.-

L., Grandone, E., Gautier, M., Bouda, S., Schloetel-

burg, C., Prinz, T., Dionisio, P., Anagnostopoulos, S.,

Drougka, I., Folkvord, F., Drosatos, G., Didaskalou,

S., and The ThrombUS+ Consortium (2024). To-

wards wearable continuous point-of-care monitoring

for deep vein thrombosis of the lower limb. In

Jarm, T.,

Smerc, R., and Mahni

c-Kalamiza, S., edi-

tors, 9th European Medical and Biological Engineer-

ing Conference, pages 326–335, Cham. Springer Na-

ture Switzerland.

Maiti, D. and Arunachalam, S. P. (2022). Non-invasive

diagnosis of deep vein thrombosis to expedite treat-

ment and prevent pulmonary embolism during gesta-

tion. volume 2022 Design of Medical Devices Con-

ference of Medical Devices, page V001T01A003.

Monteiro, E., Costa, C., and Oliveira, J. L. (2017). A

de-Identiﬁcation pipeline for ultrasound medical im-

ages in DICOM format. Journal of Medical Systems,

41(5):89.

Rodr

ıguez Gonz

alez, D., Carpenter, T., Van Hemert, J. I.,

and Wardlaw, J. (2010). An open source toolkit for

medical imaging de-identiﬁcation. European Radiol-

ogy, 20(8):1896–1904.

Waheed, S. M., Kudaravalli, P., and Hotwagner, D. T.

(2024). Deep vein thrombosis. In StatPearls. Stat-

Pearls Publishing, Treasure Island (FL).

Xiao, X., Wang, G., and Gehrke, J. (2009). Interactive

anonymization of sensitive data. In Proceedings of

the 2009 ACM SIGMOD International Conference on

Management of data, pages 1051–1054, Providence

Rhode Island USA. ACM.

Zeng, H., Li, L., Cao, Z., and Zhang, L. (2022). Grid anchor

based image cropping: A new benchmark and an ef-

ﬁcient model. IEEE Transactions on Pattern Analysis

and Machine Intelligence, 44(3):1304–1319.

HEALTHINF 2025 - 18th International Conference on Health Informatics

958