RPS 414 - AI-driven evolution: enhancing image quality, workflow, and professional identity for radiographers

Author Block: N. Stogiannos¹, G. Walsh¹, B. K. Ohene-Botwe¹, K. Mchugh², B. Potts¹, J. St John-Matthews¹, M. F. Mcentee³, Y. Kyratsis⁴, C. Malamateniou¹; ¹London/UK, ²Portsmouth/UK, ³Cork/IE, ⁴Rotterdam/NL
Purpose: Artificial intelligence is changing radiographer clinical practice and roles. It is therefore vital to understand its impact on the careers, roles and professional identity of these professionals.
Methods or Background: A European-wide, EFRS-endorsed, cross-sectional, mixed methods online survey was designed on qualtrics. Snowball sampling was used to improve uptake. Survey questions explored radiographer perceptions for the short-term and long-term impact of AI implementation on their roles, responsibilities and professional identity. The study was translated in 8 languages. Responses were compared between different demographic groups including gender, age, education and country digital literacy level.
Results or Findings: 2206 valid responses were received from 37 different countries in Europe. 50.4% reported no AI education, and 26.6% were self-taught in AI. Over half (51.1%) thought patient-centered care skills will remain the same. 50.9% agreed radiographers will have more time to spend with patients thanks to AI. 57.8% agreed radiographers will have to work closer with other MIRT professionals in the future, for efficient AI implementation. Men appeared slightly more enthused about the development of technological skills and women about the honing of patient centered care skills, similar to previous studies. Radiographers were overall optimistic about the use of AI in healthcare, and optimism was higher in those countries with high digital literacy, better education levels and with more AI experience.
Conclusion: Radiographers were overall optimistic about the use of AI in healthcare and strongly believed that AI will advance patient-centred care. AI education currently lags for European radiographers, and this should be acutely addressed at the scale and pace required to keep up with current technological developments. Interprofessional collaboration was seen as essential for fostering mutual support among professionals.
Limitations: Snowball sampling can lead to selection-bias, but allows for many recruits.
Funding for this study: Funded by the College of Radiographers Industry Partnership Scheme (CoRIPS) [grant number: 2018].
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethics approval was obtained from City St George’s, University of London School of Health and Psychological Sciences Ethics Committee (ETH2223-1346).

Author Block: L. A. Rainford, M. Mujaydia Alotaibi, J. Mcnulty, J. Potočnik; Dublin/IE
Purpose: Quality assurance (QA) of radiographic technique is an essential part of radiation protection, traditionally performed through Reject Analysis. Digital imaging has increased the difficulty in completing radiographic technique auditing and staff shortages further compromise QA monitoring. This research aimed to seek radiography opinion on the use of Artificial Intelligence (AI) for QA.
Methods or Background: An online survey was developed (n=30 questions) to seek information related to QA monitoring of radiographic technique. Participant demographics, including area of employment and country of work, and professional and AI experience, were captured. Current QA auditing details were requested and participant confidence in these processes. Their opinion was requested on the potential challenges and benefits of AI use in QA monitoring. The survey was distributed to affiliate EFRS academic institutions to distribute to their clinical training sites and via Radiography social media.
Results or Findings: Good representation across all radiography professional grades was received from 125 participants (n=22 countries). 19.8% reported QA of radiographic images on at least a weekly basis, 18.8% stated monthly, whilst 60% reported it occurred far less frequently. 20% of responses stated staff were not individually reviewed. Only 26.8% were very confident in current QA processes, 48% were somewhat confident and the remainder not confident or unsure. 80% of participants indicated they perceived AI as having a role in QA, less than 10% demonstrated concern. Improved quality standards and skills were perceived as benefits however consideration of difficult patients was an identified challenge.
Conclusion: Poor confidence in current QA processes was identified and a lack of standardisation of practice. Radiographers identified AI as having the potential to support radiographic technique audits. Benefits and challenges were identified in open comments.
Limitations: Online survey: English language could have limited uptake
Funding for this study: Self funded
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: University College Dublin, Human Research Ethics Committee – Sciences (HREC-LS) - LS-LR-24-141-Alotaibi-Rainford.

Author Block: E. Compton¹, S. Lightfoot¹, R. Shah², S. Ather², P. Taylor¹, N. H. Woznitza¹; ¹London/UK, ²Oxford/UK
Purpose: Chest radiographs (CXRs) are a high-volume test, performed for a broad spectrum of reasons. Education has been shown to improve CXR reporting accuracy, in particular for less experienced reporters. Similarly, artificial intelligence (AI) as a clinical decision support tool provides novice readers with the most benefit. The aim of this study was to compare the impact of education (SIM) with AI in CXR reporting accuracy.
Methods or Background: A multi-reader, multi-case diagnostic accuracy study was conducted to determine the impact of SIM and AI on reporting radiographer (RR) CXR accuracy. 64 RR consented and completed bank 1 and were randomised stratified by years’ experience (n=32,50% to AI). 43 RRs (24 AI, 19 education) completed both image banks (n=52 CXRs, 26 abnormal).
Results or Findings: Similar pre and post intervention accuracy was found. The AI cohort decreased sensitivity (74%-65%,p=0.015) but increased specificity (63%-77%,p<0.0001), the increase SIM sensitivity (69%-62%,p=0.115) and decrease in specificity (62%-68%,p=0.217) were not statistically significant. Standalone AI sensitivity and specificity were 54% and 77% respectively. For the AI arm, when the AI was correct specificity improved (67%-86%,p<0.001) with no significant difference in sensitivity (89%-86%,p=0.31), however when AI was incorrect there was a significant decrease in sensitivity (52%-32%,p<0.001) with no difference in specificity (both 57%,p=1). There were four CXRs that only had one (n=3) or three (n=1) pre-intervention correct decisions, suggesting the bank selected comprised of very challenging cases.
Conclusion: In a challenging CXR bank, both education and AI improved RR performance. As AI tools are adopted for CXR interpretation in clinical practice further work is required to ensure reporters are education in their use.
Limitations: The enhanced prevalence (50% abnormal) and single pathology (lung cancer) may limit transferability into clinical practice.
Funding for this study: This study was conducted as part of a clinical fellowship supported by NHS England (London).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Canterbury Christ Church University ETH2223-0246 21st April 2023

Author Block: M. R. V. Pedersen¹, M. W. Kusk², S. Lysdahlgaard², H. Mork-Knudsen³, C. Malamateniou⁴, J. Jensen⁵; ¹Vejle/DK, ²Esbjerg/DK, ³Bergen/NO, ⁴London/UK, ⁵Odense/DK
Purpose: The integration of artificial intelligence (AI) into radiography offers potential in enhancing workflow efficiency, image processing, patient positioning, and quality assurance.. This study aimed to investigate the perspectives and attitudes towards AI in radiography.
Methods or Background: An online survey including of 29 items was distributed via social media platforms to Nordic students and radiographers working in Denmark, Norway, Sweden, Iceland, Greenland, and the Faroe Islands. The survey included questions on demographics, specialization, educational background, place of work, and perspectives and knowledge on AI. The items were a mix of closed-type and scaled questions, with options for free-text responses when relevant
Results or Findings: The survey received 586 responses from all Nordic countries. The mean age was 37.2 years with a standard deviation (SD) of ±12.1 years,. A total of 43% (n = 254) of the respondents had not received any AI training in clinical practice, while 13% (n = 76) had received AI training during their radiography undergraduate studies. Additionally, 77.9% (n = 412) expressed interest in pursuing AI education. The majority of respondents (82.8%, n = 485) were aware of the potential use of AI, and 39.1% (n = 204) had no reservations about AI
Conclusion: Overall, radiographers have a positive attitude towards AI. However, there has been very limited training or education provided to radiographers, despite 82.8% reporting plans to implement AI in clinical practice. Generally, awareness of AI applications is high
Limitations: Limitations include language barriers as this survey was provided in English. Most Nordic radiographers speak, read, and write English very well. Yet, when it come to complex sentences in English there is a higher risk of skipping items, survey drop out, language misunderstanding or misinterpretation.
Funding for this study: No funding
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by the Research Ethics Committee at the University of Southern Denmark (ID: 22-58485)

Author Block: A. Gill, S. L. Mcfadden, C. Rainey, L. Mclaughlin, J. Mcconnell, C. Hughes, R. Bond; Belfast/UK
Purpose: This study investigates the cognitive behaviour of imaging staff when interacting with Explanation User Interfaces (EUI). Data was gathered on user preferences of chest radiograph Artificial Intelligence (AI)-based EUIs.
Methods or Background: Human and machine interaction involves the EUI that clinicians use to link medical diagnosis or report. However, there is currently a lack of EUI standardisation within this field (Schalekamp et al, 2022).
Building on an international questionnaire undertaken at ECR 2024, a multi-methods study was undertaken incorporating eye-tracking, Think-Aloud and a questionnaire at UKIO 2024. Diagnostic radiographers’, radiologists’, trainee radiologists’ and student radiographers’ identified visual preferences when reviewing four different types of chest radiograph AI EUIs i.e. 1) salience maps, 2) textual reports, 3) area of interest and 4)abnormality score EUIs. Participants reviewed the images whilst wearing eye-tracking software and voiced their thought processes i.e. the “Think-Aloud” method. The post study questionnaire asked the participants about their perceived level of confidence against the four different interfaces.
Results or Findings: 24 participants enabled understanding of which components of the chest radiograph EUI are focused on and subsequently preferred. Eye-tracking data relating to fixations and saccades statistically described patterns where maximal attention was directed in the interpretation process. Think-Aloud and post-study questionnaire data added further insights to participant EUI preferences. The analysis of the eye-tracking study remains ongoing, and completion is aimed for January 2025.
Conclusion: Understanding user preference for chest radiograph AI EUI supports appropriate user engagement with the information provided by the technology. This gives radiographers and radiologists the ability to explain this further to patients, positively impacting their understanding and subsequent care.
Limitations: · Small sample size may have affected the wider generalisability of findings.
· Eye-tracking software capabilities
Funding for this study: PhD funded by Department for the Economy
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: FCNUR-23-084 reference

Author Block: V. Tambè, M. Zanardo, C. Torrito, P. Della Cagnoletta, F. Secchi; Milan/IT
Purpose: Cardiovascular magnetic resonance imaging (CMR) is a critical tool for diagnosing heart disease, but is hindered by long acquisition times and manual post-processing. This systematic review examines recent advancements in image post-processing and workflow optimization, focusing on the integration of artificial intelligence (AI).
Methods or Background: A systematic search was conducted using PubMed and EMBASE. Included studies involved the use of AI-based methods to optimize CMR workflows and post-processing. Eligible articles were those addressing any of the following: image reconstruction, segmentation, workflow automation, and clinical integration of AI tools. Studies without quantitative outcomes related to workflow efficiency or post-processing improvements were excluded.
Results or Findings: Out of 151 articles screened, 33 studies were included. Key findings included: automated segmentation reported in 15/33 (45%) studies; image reconstruction in 10/33 (30%); workflow automation 8/33 (24%); clinical efficiency in 7/33 (21%); quality control in 5/33 (15%). In the automated segmentation articles, results showed improvements in segmentation speed and accuracy with Dice similarity coefficients exceeding 0.90 in many studies, and reducing manual post-processing time by up to 66%. Other studies focused on reducing scan times and enhancing image quality, with AI-based methods reducing scan times by up to 40% while maintaining image quality. Articles showed significant improvements in reporting times (by up to 30%), while 5 articles presented data on AI-based quality control reducing rescans (by up to 20%).
Conclusion: AI integration into CMR has significantly improved workflow efficiency, reducing acquisition times and enhancing diagnostic accuracy. Automated segmentation, image reconstruction, and workflow automation have accelerated processes, reduced operator dependency, and minimized rescans.
Limitations: Further large-scale validation is needed to fully implement AI in CMR across diverse clinical settings.
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: Not Applicable

Author Block: C. Schneckenleitner, C. Kamp, C. Vogl, A. Raith, G. Guevara; Vienna/AT
Purpose: The working environment of radiographers is characterized by permanent technological innovations. We developed an international intensive program containing theory and hands-on training to introduce bachelor and masters’ students to future technologies. This program is designed to prepare students for innovative technologies and expand their professional skills in areas such as deep learning models, Computer Assisted Surgery Simulation, 3D Printing, Optical Scanning and visualization utilizing mixed reality.
Methods or Background: During a supervised online phase, students learned how to create patient-specific 3D-printed models, CT-data segmentation for deep learning training with 3D-Slicer, computer-assisted surgical planning (CAS), how to acquire optical 3D-scans and mixed reality visualization of the respective 3D-models they created.
All this content was worked on by the students in a follow-up international skills lab week at the University of Applied Sciences Vienna. Each skills lab block included 6h hours of hands-on training. The students generated deep learning models with the platform MONAI, created surgical plans with Materialise Mimics, scanned with optical 3D-scanners from ARTEC and created mixed reality visualizations.
Results or Findings: The results were uploaded to an online platform (Moodle) by the students and analyzed based on defined criteria. Results, 39 out of 42 students were able to create a segmentation for deep learning training according to the required criteria. 39 out of 42 successfully performed a CAS plan, 20 created a 3D-printable patient specific model and 42 of 42 were able to produce an optical 3D-scan of the face.
Conclusion: The results show that radiography students can produce results in adjacent technology areas and expand their competence in future technologies.
Limitations: The limited time can generate interest but not a specialization in the profession. Explicit training programs are needed to deepen radiographers’ expertise in these areas.
Funding for this study: No funding.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: The results includes no patient specific information or clinical interventions.

Author Block: K. Tissir¹, G. R. Bonfitto¹, A. Roletto¹, A. Signoroni²; ¹Milan/IT, ²Brescia/IT
Purpose: The integration of Large Language Models (LLM) tools like ChatGPT in clinical settings is changing how healthcare professionals manage diagnoses and workflow. In cardiac clinics, the accurate and timely interpretation of images is crucial for effective diagnosis and monitoring of cardiac conditions. ChatGPT could be particularly beneficial for cardiac sonographers. This study aims to assess whether ChatGPT can effectively support cardiac sonographers in the qualitative evaluation of echocardiographic images.
Methods or Background: A database of 50 anonymized echocardiographic images was retrospectively analyzed, including 2-chamber, 4-chamber, and apical 3-chamber views. Three evaluators, a junior sonographer, a senior sonographer and ChatGPT-4o conducted the qualitative evaluation of the images by identifying scoring them on a 5-point Likert scale. The guidelines of the European Association of Echocardiography served as references.
Results or Findings: Junior sonographer correctly identified views in 84% of cases (n=42), while ChatGPT correctly identified 58% of cases (n=29).In comparison to senior sonographer, the junior sonographer overestimated 22% of the images (n=11), underestimated 36% (n=18), and agreed in 42% of the images (n=21). In contrast, ChatGPT overestimated 52% of the images (n=26), underestimated 18% (n=9), and agreed in 30% of the images (n=15).
Conclusion: ChatGPT-4o shows limitations in identifying echo cardiac views compared to other participants. In addition, ChatGPT is inclined to overestimate image quality. This can be explained by limited training of the LLM, mainly done with information from guidelines. As other studies in literature showed, more in-depth training could increase the performance of ChatGPT. LLM can assist cardiac sonographers in qualitative analysis of images and supporting anomaly evaluation, but concerns remain about its reliability and bias.
Limitations: The small sample of participants and cases limited the strength of the conclusions of this study.
Funding for this study: N/A
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: University of Brescia

Author Block: G. Walsh¹, Y. Kyratsis², A. Goodall¹, J. St John-Matthews¹, C. Malamateniou¹; ¹London/UK, ²Rotterdam/NL
Purpose: This study offers unique insights into the preparedness of radiographers to pursue AI leadership roles within healthcare and potential barriers preventing radiographers excelling in the AI ecosystem.
Methods or Background: A European-wide, cross-sectional study utilising a mixed methods online survey. Snowball sampling allowed qualified radiographers, and radiography students, to answer the survey, irrespective of their current role. The survey explored the following areas of interest: a) general radiographer demographics, b) radiographers preparedness and confidence to lead the implementation of AI in healthcare, c) suggested day-to-day responsibilities of an AI-lead radiographer and d) motivations for considering AI leadership roles.
Results or Findings: There were 1733 valid responses from 37 European countries. The typical respondent was female (64%), a diagnostic radiographer (59.9%) with >20 years’ experience (31.3%). A lack of education, training and resources led 72.3% of radiographers to feel the profession is not prepared to lead the implementation of AI in healthcare. Those that felt confident to lead in an AI-enabled work environment (50.06%), felt they already have the necessary experience and skills, but also admitted additional resources would be needed. The top two motivators to pursue an AI leadership role included championing change and the promise of appropriate training.
Inferential Statistics ongoing as of October 2024.
Conclusion: Radiographers have a unique skill-set making them the ideal candidates for AI leadership roles within healthcare. Radiographers do not currently feel confident or prepared to undertake AI leadership roles with education, training and a lack of resources creating barriers for this. It is reassuring radiographers feel motivated to undertake AI leadership roles, however increased training and educational support are needed.
Limitations: Gives a snapshot view of radiographers perceptions.
Snowball sampling can lead to selection-bias, but allows for many recruits.
Funding for this study: This research has been funded by the AI special call of the College of Radiographers Industry Partnership Scheme (CORIPS) of the College of Radiographers (Reference Number 218).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethics approval (ETH2223-1346) was granted by City, University of London.

Author Block: K. M. Mead, T. Cross, G. Roger, R. Sabharwal, S. Singh, N. Giannotti; Sydney/AU
Purpose: Several studies have demonstrated that deep learning (DL) models can be effectively trained on MRI data to assist clinicians in identifying knee injuries and pathologies. This systematic review was conducted to explore the current landscape of existing DL models developed for detecting knee injuries and pathologies through magnetic resonance imaging (MRI) and to assess their potential clinical applications.
Methods or Background: Five databases were systematically searched using the following terms ‘Knee AND 3D AND MRI AND Deep Learning’. The Covidence platform was used to screen publications based on title, abstract, and full text. Only original research articles focussing on the automatic detection of knee injuries and pathologies using DL models for MRI were included. The synthesis of results was performed by two independent reviewers.
Results or Findings: Fifty-four studies were included. The studies focused on anterior cruciate ligament injuries (n=19), osteoarthritis (n=9), meniscal injuries (n=13), general abnormal knee appearance (n=10), tibial plateau fractures (n=1) and synovial fluid detection (n=1). The following convolutional neural network (CNN) infrastructures were used: ResNet, VGG, DenseNet, and DarkNet. The averaged performance outcomes of the DL models demonstrated sensitivity, specificity, AUC-ROC, and accuracy of 87%, 90%, 92%, and 88%, respectively. The DL models for the detection of a specific injury or pathology outperformed those for general abnormality detection.
Conclusion: This systematic review underscores that fine-tuned DL models for knee pathologies and injuries can effectively support automatic diagnosis. Further large-scale validation and prospective studies are needed to confirm their clinical utility as assistive diagnostic tools.
Limitations: Inconsistent data reporting across the studies analysed resulted in variations in the reporting of DL model performance. Sub-group analyses were performed to minimise bias.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: The study is retrospective. Ethics approval was deemed unnecessary by the Research Integrity and Ethics Committee at the University of Sydney, Australia.

Author Block: R. S. T. Ribeiro, T. Coutaudier, L. Mourot, C. S. D. Reis, L. Raileanu; Lausanne/CH
Purpose: To enhance the learning process of chest X-ray/(CXR) anatomy for medical imaging/(MI) students by integrating AI segmentation and classification tools into an educational web-application/(webapp).
Methods or Background: The webapp was designed as an interactive platform where students can identify CXR anatomy. The platform was implemented in Python using Flask to provide the web-interface, TorchXRayVision to classify and segment key regions (heart, lungs, clavicles, spine, scapula, trachea) with AI. PostgreSQL stored and managedCXR public datasets allowing students’ practice. The user interface allows selection and outline regions of interest on the radiographs, that are compared to segmentations obtained with AI-algorithms. Feedback is provided through Dice coefficient that assess the accuracy of the user’s segmentation compared to the AI-based reference. The app’s system architecture is modular.
Results or Findings: The developed framework successfully integrates AI for fast and accurate segmentation of CXR. Its design allows users to upload their own radiographs or use others supplied by public datasets, interact with radiographs by selecting anatomical regions and receiving immediate feedback. This functionality aims to support autonomous learning and reduce the need for constant instructor supervision. The modular architecture ensures scalability, enabling the inclusion of more types of radiographic images and enhancing its potential for broader applications in medical imaging education.
Conclusion: The web application demonstrates a promising approach to improve MI education by providing an interactive and AI-powered learning tool. Its design is intended to be adaptable/accessible through any web-browser, with potential to expand into areas such as quality assessment. Further development of this framework is planned to test its impact on the MI students learning process.
Limitations: The app is limited to CXR and relies on AI segmentation performance.
Funding for this study: Not applicable
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: Not applicable