RPS 2205 - Imaging informatics, quality and new techniques

Author Block: B. O. Klaudel¹, A. Obuchowski², A. Komor¹, P. Frąckowski¹, K. Rogala¹, K. Knitter¹; ¹Gdańsk/PL, ²Banino/PL
Purpose: We present a new standard for creating and unifying large-scale diagnostic imaging datasets. It aims to address the challenges in medical AI, particularly the lack of standardized, comprehensive data for training foundation models in medical imaging. By combining multiple open-source datasets, unifying them to a common ontology and providing standardized preprocessing pipelines, we seek to accelerate the development of more generalized and robust medical AI models.
Methods or Background: We combined over 20 open-source datasets, resulting in more than one million annotated medical images, including CT, MRI, and X-ray modalities. We created reusable preprocessing pipelines to transform diverse source datasets into a unified format. A key innovation is the adoption of the RadLex ontology, developed by the Radiological Society of North America, to standardize labels and masks across all included datasets. It addresses issues of inconsistent labeling, regional variations, and differing granularity in existing medical imaging datasets.
Results or Findings: The project has created the largest publicly available dataset of annotated radiological imaging to date, with over one million images, 40+ labels, and 15 annotation masks. The datasets are accompanied by ready-to-use preprocessing pipelines that can be easily adapted to incorporate new data sources. The unified ontology based on RadLex enables consistent labeling, facilitating more effective model training and cross-dataset compatibility.
Conclusion: UMIE datasets represent a significant advancement in medical imaging AI, providing a standardized foundation for developing more generalized models. By addressing the challenges of data scarcity, inconsistent formatting, and labeling discrepancies, it paves the way for more robust and widely applicable AI solutions The open-source nature of the project encourages collaboration and further expansion of the datasets.
Limitations: UMIE datasets inherits potential biases and incomplete labelling coverage from source datasets, and limitations of the RadLex ontology.
Funding for this study: No funding was required.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: All data used in this study come from existing opensource datasets.

Author Block: P. Pilius¹, N. Smirnov²; ¹Almaty/KZ, ²Haar/DE
Purpose: The aim of this work is to analyze the fundamental principles of project management and evaluate potential opportunities for its implementation for improving efficiency in the preparation of medical data for machine learning: for data collection, working with annotators, and results validation.
Methods or Background: We analyzed personal experiences from 9 projects conducted during 2022-2024 years, aimed at creating annotated medical datasets with various levels of complexity and across different fields. The clients were private companies developing artificial intelligence technologies, and the annotators included certified radiologists, residents, and radiographers. To enhance efficiency, we studied and applied principles from an advanced training course on the foundations of project management.
Results or Findings: The duration of the projects ranged from 3 weeks to 3 months. The number of annotated datasets included: 75 MRI, 3720 CT, 55,500 X-rays, 6000 angiograms, 1700 mammograms, and 3200 radiological text report annotations. All projects were completed on time and met quality requirements. Based on practical experience and the implementation of fundamental project management concepts, we've formulated recommendations and standards for conducting similar projects.
Conclusion: The organization of medical data annotation processes for machine learning presents numerous specific challenges. To optimize the workflow and achieve desired outcomes, it is essential to have not only sufficient experience in handling medical images but also a solid understanding of fundamental project management principles.
Limitations: The work is based on personal experience and observations and is primarily of a recommendatory nature. It is not possible to objectively compare the outcomes of projects where basic management principles were and were not applied, as the projects were not repeated under identical conditions. However, the variability and uniqueness of each project allowed for a comprehensive analysis, leading to the formulation of universal standards.
Funding for this study: No funding
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Kazakh National Scientific and Research Center of Oncology and Radiology

Author Block: K. Nairz, N. Cihoric, F. Dennstädt, M. Schmerder, H. Bonel, H. Von Tengg-Kobligk; Bern/CH
Purpose: Structured reporting (SR) in radiology has been shown to be the most favorable scheme to provide imaging information to referrers. Implementation of SR is associated with additional effort for the radiologists, but there are recent advances with Large Language Models (LLMs) that are prompted to convert free text into a structured format. We aimed at killing two birds with a stone by leveraging LLMs to generate structured data that enhances interoperability as well. As a proof of concept we we selected breast cancer patient pathways for documentation. We structured the free-text information from various sources, including health interviews, mammography reports, biopsy results, pathology findings, and tumor board discussions, to ensure that the data could be effectively transmitted and support therapeutic decision-making.
Methods or Background: Our approach is based on the use of Common Data Elements (CDEs), which are minimal information units or precisely defined questions associated with a set of standardized answers, each having explicitly defined values. Building on in-depth analyses of reports and guidelines such as BI-RADS, we defined corresponding sets of Common Data Elements and created templates to establish a structured format. To ensure data protection we utilized a locally installed LLM (Llama 3), which was prompted to answer CDE-based questions. This process enabled the mapping of free-text content to a structured format, which was ultimately stored in FHIR compatible JSON format.
Results or Findings: By defining key CDE values, or specific combinations of these values, clinicians can identify critical insights that may suggest a particular therapeutic course or provide predictive indicators for patient outcomes.
Conclusion: Prompting LLMs to answer CDE-based structures proves to be a viable approach to promote data interoperability in complex medical settings.
Limitations: The study focuses specifically on breast cancer pathways.
Funding for this study: Innosuisse 59228.1
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Kantonale Ethikkommission Bern

Author Block: A. S. Quinsten¹, A. Wetter², M. Raczkowski³, L. Trembecki³, R. Buchkremer⁴, D. Matusiewicz¹, K. Nassenstein¹, M. Forsting¹, A. Demircioglu¹; ¹Essen/DE, ²Hamburg/DE, ³Wrocław/PL, ⁴Düsseldorf/DE
Purpose: Prostate magnetic resonance imaging (MRI) is typically conducted according to manual prescriptions by radiographers. This approach is time-consuming, error-prone, inconsistent due to rater variability, and has low reproducibility. The aim of the study was to develop a deep learning-based framework for the automatic planning of the field of view (FoV) in the oblique coronal and axial planes in prostate MRI according to Prostate Imaging Reporting and Data System (PI-RADS) guidelines.
Methods or Background: The retrospective multicentre study included 2109 patients from diagnostic (Sites I and III) and radiotherapy (Site II) centres. The variability within and between raters was evaluated by three assessors. Three distinct deep neural networks were developed with the objective of predicting the oblique coronal and axial FoV. The optimal network was evaluated on three external cohorts using a non-inferiority test, and its clinical utility was assessed.
Results or Findings: The optimal model demonstrated non-inferior performance, with slice position differences ranging from 0.21 ± 0.99 and 0.37 ± 0.48. At Sites I and III, the predictions were predominantly non-inferior, with FoV overlaps of 86.6 ± 5.8% and 88.7 ± 6.0% and angle differences of 4.66 ± 4.89° (Site I) as well as 3.46 ± 2.80° (Site III). In contrast, the predictions for Site II demonstrated inferior overlap (67.0 ± 9.7% and 63.6 ± 8.8%) and higher angle differences (9.18 ± 9.49°). Consequently, the clinical utility was excellent for Sites I and III (97.9–100%) but lower for Site II (85.3–89.0%).
Conclusion: The utilisation of a deep learning-based framework for the automated positioning of the FoV in oblique coronal and axial planes for prostate MRI is a viable approach, exhibiting high clinical utility.
Limitations: The present study did not include images acquired with the endorectal coil.
Funding for this study: No funding was provided for this study
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The ethics committee notification can be found under the number 22-10740-BO.

Author Block: C. S. Schmidt, M. Walter, J. Haubold, F. Nensa, R. Hosch; Essen/DE
Purpose: The aim of this study was to develop a deep learning (DL) model for the automatic detection and localisation of medical devices known to cause metal artefacts in CT images, utilising their corresponding topograms.
Methods or Background: A dataset of 943 CT topograms with radiopaque medical devices was manually annotated via box labelling by a radiology resident with three years of experience in CT imaging. The following classes were defined: cochlear implant, cardiac conduction device (pacemaker, defibrillator, stimulator), implanted port, prosthetic heart valve, (embolisation) coil, osteosynthesis (nail-, plate-, screw-, and wire-fixation, spinal instrumentation hardware), sternal wires, external fixation hardware, hip prosthesis, shoulder prosthesis, knee prosthesis, denture (prosthesis, implant). An 80/10/10% split for training, validation and testing was performed and the YOLO11X model was trained for 100 epochs. The model was evaluated using mAP50 scores, precision (P) and recall (R).
Results or Findings: The model achieved an average mAP50 score of 0.83, Precision of 0.86 and Recall of 0.79 over all classes and the following (mAP50/P/R) scores for the respective classes: cochlear implant (0.92/0.93/0.85), cardiac conduction device (0.90/0.82/0.93), implanted port (0.91/0.97/0.79), prosthetic heart valve (0.86/0.82/0.78), coil (0.99/0.98/1), osteosynthesis (0.63/1/0.55), sternal wires (0.51/0.74/0.57), external fixation hardware (0.52/0.51/0.33), hip prosthesis (0.99/0.89/1), shoulder prosthesis (0.88/0.88/0.86), knee prosthesis (0.99/0.95/1), denture (0.88/0.81/0.76).
Conclusion: The presented model demonstrates an accurate detection of most radiopaque medical devices in CT scout images. It could thus be utilised as an efficient orchestration tool for selecting a cohort of high quality imaging studies without interfering artefacts.
Limitations: The limitations of the study are its small sample size and that scout images were annotated by a single observer. Additionally, certain medical devices can be challenging to identify and localise on topograms, which could cause relevant features to go undetected.
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: Informed consent was waived by the ethics committee due to the retrospective setting.

Author Block: F. Mariotti, A. Agostini, A. Borgheresi, L. Pierpaoli, F. Ricciardiello, A. Zannotti, D. Nicolini, A. B. Galosi, A. Giovagnoni; Ancona/IT
Purpose: This study aims to evaluate the impact of quantum technology on the performance of radiomics random forest (RF) models for medical imaging. We simulated a semi-quantum approach, involving quantum embedding followed by classical RF, and a fully quantum approach using a quantum random forest (QRF) model.
Methods or Background: We used three radiomic datasets: 1. Perineural infiltration of peripancreatic fat in pancreatic adenocarcinoma on CT, 2. Characterization of renal nodules in CT, 3. Prediction of LI-RADS category on abbreviated MRI protocols. For the quantum approaches, we compared the original random forest (RF) models with simulated quantum-embedded RF and quantum random forest (QRF) algorithms, implemented in Python using an 8-qubit configuration. The comparison involved analyzing the accuracy and the receiver operating characteristic (ROC) curves using statistical significance set at p-values < 0,05
Results or Findings: The classical RF achieved the highest accuracy for the pancreas (0.9167) and kidney (0.8571) datasets. For the liver dataset, both the quantum embedding RF and QRF outperformed the classical approach (0.8462 vs. 0.7692), with the ROC curves showing statistically significant improvement (p < 0.01). In the pancreas dataset, quantum methods showed slightly lower accuracy (0.8333), and for the kidney, they also performed worse (0.7857). This indicates that the benefits of quantum approaches may be data-dependent, providing advantages in some cases but not yielding consistent improvements across all datasets.
Conclusion: Quantum machine learning is a feasible approach for radiomic datasets, showing variable results and the potential to outperform classical methods. However, the variability in performance suggests that fine-tuning of quantum algorithms may be necessary depending on the specific characteristics of each dataset.
Limitations: Small datasets used and simulation of quantum processes with a 8-qubit setup. Further research should involve larger datasets and physical quantum devices.
Funding for this study: This study did not receive any specific funding from public, commercial, or not-for-profit sectors. The research was conducted without external financial support.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: Not Applicable

Author Block: A. M. C. Boehner, A. Jacob, A. Isaak, C. C. Pieper, J. A. Luetkens, D. Kütting; Bonn/DE
Purpose: 3D data is rarely spatially displayed in routine. However, patient-clinician and clinician-clinician interaction may benefit from such representation in Extended Reality (XR). Additionally, radiologists may aid surgeons during surgery via audiovisual communication to demonstrate 3D data if needed.
Methods or Background: We developed and tested a workflow integrating different software platforms (e.g.‘Medical Imaging XR’, ‘Fiji’) to display DICOM images on iPhone and iPad (n=35) and Apple Vision Pro (AVP, n=10). The system enables fused XR visualization of CT, MRI, PET. Handheld devices were utilized to aid sonographic correlations of hepatic lesions (n=11); by surgeons during surgery preparations (n=10); and for patient information (n=14). Integrated systems were tested in a mock audiovisual call from the operating room via the AVP to the other devices located on and off campus.
Results or Findings: Our framework allowed for fast integration of 3D datasets across devices with low computational burden. XR during sonographic correlation of hepatic lesions significantly reduced the time needed to identify lesions from 4:50min to 2:45min (P<0.05). Patient reported full acceptance of XR usage. AVP allowed real-time image-data and view sharing between the radiologist and surgeon.
Conclusion: Integration of XR across smartphones, tablets and AVP enhanced medical imaging communication between all parties, reducing time to locate lesions and improving patient-physician interactions. AVP further facilitates sterile audiovisual communication between surgeons and radiologists during procedures, allowing for remote and swift consultation without leaving the sterile field.
Limitations: Our method was tested exclusively on Apple products, limiting its generalizability to other platforms.
Funding for this study: This project was part of the ISMC, funded by the Ministry of Economic Affairs, Innovation, Digitalization and Energy of the state of North Rhine-Westphalia
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethics University Hospital Bonn, Germany (2024-228-BO)