PIXEL PANDEMONIUM
European Congress of Radiology 2026Pixel Pandemonium at ECR 2026
Location: Foyer K (Level -2) Dates: March 4–7, 2026
Pixel Pandemonium showcases cutting-edge AI tools, platforms, and prototypes developed by researchers and innovators from across the medical imaging community. Visit Foyer K (Level -2) to explore live demos and interact directly with the developers behind these projects.
Schedule Overview
Pixel Pandemonium features 10 exhibition booths. Some exhibitors are present throughout the entire congress, while others rotate between two time slots:
- Full Slot (All Days): Wednesday, March 4 – Saturday, March 7
- Slot A: Wednesday, March 4 – Thursday, March 5
- Slot B: Friday, March 6 – Saturday, March 7
An overview of all exhibitions with details on the demos is included below. Be sure to stop by and experience these previews hands-on at ECR 2026!
Full Slot Exhibitors (All Days: March 4–7)
Booth 1 — FairGrowth: An Interactive Tool to Uncover Bias in Deep Learning Fetal Growth Models
Aya Elgebaly · Copenhagen, Denmark
Artificial intelligence models for fetal growth assessment are increasingly evaluated using global performance metrics such as AUC or sensitivity. However, aggregated metrics may obscure clinically meaningful disparities across demographic, biological, and technical subgroups.
FairGrowth is an interactive bias auditing framework designed to systematically evaluate fairness in fetal growth prediction models. The platform enables subgroup-level performance analysis across multiple dimensions, including maternal ethnicity, BMI, gestational age, hospital site, ultrasound device, parity, smoking status, hormonal stimulation, pixel spacing, and fetal sex.
FairGrowth supports comparison of three modeling approaches: a deep learning model, a DINO-based vision transformer, and the Hadlock clinical formula. Bias is quantified using sensitivity gaps between advantaged and disadvantaged subgroups, complemented by bootstrap-based uncertainty estimation. Intersectional analysis allows evaluation of compounded disparities across combined factors.
Applied to a multi-center cohort for SGA and LGA detection, the framework reveals performance heterogeneity linked to site, acquisition parameters, and patient characteristics. FairGrowth provides a transparent and reproducible approach for auditing AI systems before clinical deployment, promoting equitable model evaluation in perinatal imaging.
Booth 2 — radGPT: Instant Voice to Structured Report
Philipp Arnold · Freiburg, Germany
radGPT is a voice-to-structured-report solution designed specifically for radiologists. It transforms natural clinical dictation into fully structured, standardized radiology reports within seconds. Instead of forcing radiologists to navigate rigid templates or manually format free text, radGPT allows them to dictate freely while a large language model handles the rest. The result is a clean, guideline-conform report that is ready for direct use.
The workflow is simple: select an examination template, dictate naturally, generate the structured report, and review the built-in consistency check. radGPT automatically verifies internal logic and detects common issues such as laterality mismatches or contradictions before the report is finalized. This reduces avoidable errors while preserving diagnostic flow.
The platform runs as a browser-based web application with no installation or system integration required. Reports are generated as copy-ready text compatible with any existing RIS environment. radGPT supports multilingual reporting and is optimized for high-volume modalities such as CT and MRI.
Developed by radiologists for radiologists, radGPT focuses on speed, clarity, and reliability. It removes administrative formatting work and enables structured reporting that feels effortless in everyday clinical practice.
Co-authors: Elmar Kotter, Yassine Aguilo, Maurice Henkel, Johannes Jahn
Links: radgpt.io
Related session: RPS 2005 — How Large Language Models Are Transforming Radiological Reporting (Saturday, March 7, 14:00–15:30)
Booth 4 — mAIstro: A Multi-Agent Framework for Autonomous AI Development in Medical Imaging
Eleftherios Tzanis · Heraklion, Greece
mAIstro is an open-source, multi-agent system for end-to-end AI development in medical imaging and tabular health data. A master agent coordinates eight task-specific agents responsible for exploratory data analysis (EDA), feature importance analysis, radiomic feature extraction, medical image segmentation, classification and regression modeling. Using natural language prompts, the system interprets user requests, selects appropriate tools, and executes multi-step workflows.
The framework was evaluated on 20 public datasets, including structured clinical data and multimodal imaging datasets. A diverse set of single- and multi-task queries assessed correct agent selection, tool execution, and sequential coordination. Outputs were verified through log inspection and independent re-execution of all tools. Performance was further tested across multiple large language models (LLMs).
High-performing LLMs achieved 100% task success rates. The system autonomously generated EDA reports, feature rankings, blended predictive models, nnU-Net-based segmentations, radiomic feature sets, and deep learning image classifiers. Complex workflows, including segmentation-to-radiomics-to-classification pipelines, were completed using only natural language instructions.
mAIstro provides an LLM-agnostic framework for autonomous medical AI development, enabling both non-programmers and advanced users to build, evaluate, and extend AI pipelines using natural language interaction.
Co-authors: Michail E. Klontzas
Links: GitHub · Publication
Booth 5 — Grand Challenge: The AI Validation Platform for Radiology
Miriam Groeneveld · Nijmegen, Netherlands
AI in radiology is advancing rapidly, but proper validation and clinical oversight remain essential and underexplored. Grand Challenge enables radiologists to actively shape AI by annotating data, validating algorithms, running reader studies, and defining the benchmarks that determine clinical quality.
The platform supports the collection of annotated data to be used as training data for AI development or as the gold standard for AI benchmarking. Hosting AI models on the platform allows radiologists to test and explore these models using their own data. Benchmarking AI models against the gold standard set by radiologists builds trust and offers fair and transparent comparisons.
At ECR, the Grand Challenge team invites you to experience their Reader Study service — a powerful solution designed to support clinical validation and expert annotation workflows. The Reader Study environment enables you to run structured multi-reader studies, collect high-quality expert annotations and metadata, assess inter-reader variability and expertise, and set the gold standard for AI evaluation and benchmarking.
Join the live demo and discover how Grand Challenge helps radiologists take control of AI validation and shape the future of medical imaging.
Co-authors: James Meakin, Paul Gerke, Harm van Zeeland, Anne Mickan, Chris van Run, Thomas Koopman
Links: grand-challenge.org
Booth 8 — Eyes on the Image, Not the Keyboard: A Pragmatic AI Reporting Workflow for Radiology
Oleksander Berezovskyi · Odesa, Ukraine
In many low-resource healthcare settings, commercial medical speech-to-text (STT) systems are inaccessible. This demo presents a clinician-built, modular AI workflow designed to reduce cognitive load and eliminate focus-switching between imaging workstations and text editors.
A local-first architecture uses MacWhisper (OpenAI Whisper "large" model) for real-time STT on an M2 Max workstation. Recognized text is processed via modality-specific prompts and routed to either a cloud-based LLM (GPT-4o-mini via OpenRouter) or a fully local environment (Ollama with MedGemma 27B). To ensure privacy, no protected health information (PHI) is transmitted; all patient identifiers are handled locally.
A 12-month clinical implementation showed report generation latency of 10 seconds (cloud) vs. 30–50 seconds (local). Cloud cost averaged approximately $0.002 per report. Efficiency gains reached a 30% reduction in reporting time for complex oncologic MRI and up to 50% for routine musculoskeletal studies.
Modular, low-cost AI solutions bridge the gap between commercial availability and clinical necessity, enhancing productivity while maintaining data control in resource-limited radiology settings.
Links: MacWhisper · Ollama · MedGemma · OpenRouter
Slot A Exhibitors (Wednesday, March 4 – Thursday, March 5)
Booth 3 — Hope4kids: The Pediatric Brain Tumor Segmenter
Daniel Capellán-Martín · Madrid, Spain
Pediatric CNS tumors are the leading cause of cancer death in children. Although pediatric high-grade brain tumors are rare, they can be highly aggressive — with pediatric diffuse mid-line gliomas (DMGs) showing a median overall survival of less than one year and high-grade gliomas demonstrating under 20% 5-year survival — often compounded by delayed diagnosis. Specialized imaging tools for pediatric brain tumor analysis are crucial to improving clinical management. Automatic tumor segmentation is a vital first step for accurate quantitative analysis that supports clinical trials and personalized care.
Hope4kids is an AI algorithm designed for automatic segmentation of pediatric brain tumor structures. By processing four standardized MRI sequences (native T1, contrast-enhanced T1, T2, and T2-FLAIR), the tool generates highly accurate segmentation masks and outputs in NIfTI format. Beyond pediatric brain tumors, the platform also supports the segmentation of gliomas, brain metastases, and intracranial meningiomas. With its user-friendly interface for non-deep learning experts, Hope4kids significantly reduces manual annotation workload while facilitating quantitative analysis and supporting clinical decision-making.
Co-authors: Abhijeet Parida, Zhifan Jiang, Nishad Kulkarni, Krithika Iyer, Austin Tapp, Syed Muhammad Anwar, María J. Ledesma-Carbayo, Marius George Linguraru
Links: segmenter.hope4kids.io
Booth 6 — Deep Learning for Fibrotic and Hypoattenuation Pattern Quantification in Chest CT
David Montalvo-Garcia · Madrid, Spain
This demo presents a deep learning-based algorithm for segmenting radiological patterns in chest CT scans of patients with pulmonary fibrosis. The tool enables the identification and quantification of fibrotic patterns such as ground glass and fibrotic lesions, as well as emphysematous and air-trapping areas (jointly identified as hypoattenuation areas). The model has been trained using weak labels on state-of-the-art segmentation networks.
By supporting improved clinical understanding and quantification of pulmonary fibrosis patterns, this tool offers practical utility for both research and clinical assessment of fibrotic lung disease.
Booth 7 — Spatio-Temporal Deep Learning for Lung Nodule Malignancy Prediction
Maria J. Ledesma-Carbayo · Madrid, Spain
This demo predicts the malignancy probability of lung nodules by analyzing up to three 3D CT scans taken at different timepoints. Using spatio-temporal deep learning, the tool evaluates the temporal progression of nodules to provide a malignancy probability, attention weights to highlight the importance of each timepoint, and saliency maps for interpretability. Users can input complete or partial data, making the tool versatile for exploring nodule behavior in clinical and research settings.
Links: GitHub
Booth 9 — LLM Extractinator: A Toolkit for Extracting Structured Medical Data from Text
Luc Builtjes · Nijmegen, Netherlands
LLM Extractinator is an open-source, modular toolkit for converting free-text clinical documents into structured datasets using large language models. Clinical narratives such as radiology and pathology reports contain valuable information, but manual abstraction is labor-intensive and traditional NLP pipelines are difficult to adapt across tasks and institutions.
The framework uses a schema-first approach with prompt templates, few-shot examples, and constrained generation to produce validated JSON outputs aligned with user-defined schemas. A simple graphical user interface allows users to configure and run extraction workflows without coding knowledge. The system runs entirely with locally hosted models and in reproducible, containerized environments suitable for hospital IT settings. By standardizing LLM-based information extraction while remaining transparent and easy to use, LLM Extractinator lowers the barrier to structured clinical data curation.
Co-authors: Joeran Bosma, Mathias Prokop, Bram van Ginneken, Alessa Hering
Booth 10 — AI-Based Opportunistic Cardiovascular Screening on Breast MRI
Dimitrios Bounias · Heidelberg, Germany
This demo presents an AI tool for fully automated secondary health assessment on routine breast MRI examinations. The system integrates with the hospital PACS, automatically retrieves examinations, analyzes them in the background, and returns a structured, easy-to-read report to the radiologist — without changing the existing imaging workflow.
The tool focuses on clinically relevant incidental findings of the central cardiovascular system that may be overlooked during routine breast MRI readings, including aortic aneurysms, pulmonary artery dilation, and cardiomegaly. The report summarizes key measurements alongside reference values and provides visual examples of the measured structures to support fast review. By opportunistically using already acquired imaging data, this work demonstrates how background AI analysis can improve patient safety and support radiologists without requiring additional examinations.
Slot B Exhibitors (Friday, March 6 – Saturday, March 7)
Booth 3 — DentiHUB: Transforming Digital Dentistry Education and AI-Driven Radiodiagnostics
Hedesiu Mihaela · Cluj-Napoca, Romania
DentiHUB is a digital educational platform designed to advance dentists' competencies in digital treatment planning for maxillofacial surgery. By facilitating structured exposure to digital workflows, the platform enhances compliance in the design and implementation of surgical guides and supports the transition toward personalized dental medicine.
DentiHUB integrates a comprehensive imaging database, including CT/MRI, CBCT, 2D dental radiographs, intraoperative images, as well as surgical plans and guides. In addition, the platform features a dedicated research component that leverages artificial intelligence applications in 2D and 3D CBCT dental radiodiagnostics, bridging clinical practice with data-driven innovation. The platform delivers structured educational content, including video-based learning materials and courses focused on digital workflows in dentistry, supporting the adoption of evidence-based, technology-driven clinical practice.
Links: dentihub.ro · DentiHUB Platform
Booth 6 — Standardizing the Evaluation of Foundation Models in Radiology: The Healthcare AI Challenge within the MGB AI Arena
Marcio Aloisio Bezerra Cavalcanti Rockenbach · Somerville, United States
The Mass General Brigham AI Arena tackles a major gap in healthcare AI: the lack of a standardized way to evaluate foundation models in real clinical settings, especially medical imaging. The Arena provides secure infrastructure, access to real-world datasets, and scenario-based testing that mirrors clinical workflows, including draft radiology report generation.
The centerpiece of this effort is the Healthcare AI Challenge, a flagship program that turns the Arena into a multi-institution collaboration. Through the Challenge, clinicians and domain experts test and compare AI solutions in realistic clinical scenarios. Participants provide structured feedback that generates actionable insights for health systems, researchers, and model developers — moving beyond isolated performance metrics toward evidence of clinical utility, limitations, and operational readiness.
Looking ahead, the evaluation framework will go beyond accuracy alone, building methods to quantify efficiency gains as agentic workflows become more common in reporting tasks and other clinical operations. The goal is to identify AI solutions that augment clinical teams in ways that are safe and measurable.
Co-authors: Sarah Mercaldo, Laura Brink, Alexander Schultz, Ben Lewis, Ryan Morley, Thomas Pryor, Laura Coombs, Richard Bruce, Dushyant Sahani, Nabile Safdar, Christoph Wald, Thomas Schultz, Keith Dreyer, Bernardo Bizzo
Booth 7 — ATLAS: The Annotated Library of AI Systems
Charles Kahn · Philadelphia, United States
ATLAS, the Annotated Library of AI Systems, indexes information about AI models and datasets to make them more transparent and discoverable. Each ATLAS "card" uses terms from the ROADMAP ontology to describe multi-modal AI models and datasets; cards are indexed by subspecialty, RadLex terms, and keywords. Users can browse ATLAS cards on the web or using an API.
Booth 9 — CuraForge: Empowering Users to Transform AI Results and Clinical Data into Report-Ready Text
Joshy Cyriac · Basel, Switzerland
CuraForge Engine is a modular integration platform that unifies external AI solutions and clinical data services such as RIS, DICOM, and HIS within one radiology workflow. Its integrated Python scripting engine transforms structured data and AI outputs into automated text elements that are directly embedded into reporting templates.
Radiology reporting often suffers from fragmented AI results and rigid systems. CuraForge consolidates diverse data sources, enables custom data transformation, and injects processed insights seamlessly into the reporting environment. Core features include automated reporting of AI findings and technical parameters, an LLM-powered quality control module for laterality and consistency checks, and an ICD-10 coding assistant.
The system automatically generates text for radiation dose and contrast information, prior comparisons, gender-specific statements, MRI machine details, and AI findings for lung, pleura, pericardium, and aorta. Failed quality checks trigger automatic email notifications to the reporting radiologist.
In daily production since January 2025 for all CT and MRI body and chest exams, CuraForge reduced dose, contrast, and gender-related reporting errors to zero. It saves 666.5 reporting hours annually — approximately 2% of report writing time for 31 residents — freeing approximately USD 76,742 in labor costs.
Links: curaforge.io
Booth 10 — DeepTwin-X: An Interactive GenAI Virtual Scanner for Multimodal Radiological Digital Twins
Francesco Di Feola · Umeå, Sweden
DeepTwin-X is an interactive prototype that lets clinicians and researchers generate missing imaging modalities in real time, bridging the gap between generative AI research and clinical usability. It contributes to the broader vision of a radiological Digital Twin: a virtual patient designed to support understanding, simulation, and clinical decision-making.
In healthcare, data are multimodal but often incomplete, and current AI systems typically analyze only one modality at a time. Generative AI offers new possibilities to synthesize missing information and integrate diverse data sources into coherent patient models. The framework is structured around three components: virtual scanner, virtual treatment, and decision support — together laying the groundwork for data-driven clinical decision-making.
This demo focuses on the virtual scanner and presents three translation use cases: MRI-to-CT, CBCT-to-CT, and CT-to-PET, currently under evaluation through expert assessment and quantitative benchmarking. Users can interact directly with the interface: uploading imaging inputs, running live modality translations, and comparing synthetic outputs side by side. By embedding these models in an accessible, user-facing interface, DeepTwin-X demonstrates that Digital Twin technologies can be both algorithmically powerful and suited for real-world workflows.
Co-authors: Valerio Guarrasi, Giulia Romoli, Filippo Ruffini, Paolo Soda