Challenges

The ISBI 2026 Challenges aim to accelerate innovation in biomedical imaging by promoting the development, evaluation, and benchmarking of cutting-edge AI and computational imaging methods. Participants will tackle a diverse set of problems across ultrasound, X-ray, magnetic particle imaging, endoscopy, cytology, and blood analysis, addressing clinically and biologically significant challenges. Each challenge provides a platform to advance reproducible, generalizable, and clinically impactful algorithms, fostering collaboration and competition in the scientific community.

Imortant Dates

• February 26  – Paper Submission Deadline
  • Challenge participants submit their papers on the EDAS platform (challenge track)
  • Review period is open for 2 weeks
  • Reviewers are assigned by the organizers of each challenge
  • Matareview is carried out by the organizers of each challenge
  • Method evaluations are still ongoing (i.e. challenge is still running)
• March 1- Challenge Paper Authors Receive Reviews Through the EDAS Platform
  • We recommend to organizers to stop accepting method evaluations
  • Organizers of each challenge should notify challenge paper authors that they have 1 week to address comments and to add final methodological changes
  • Organizers of each challenge invite every valid* paper submitter to
    • Present their method at the conference (oral or poster will be decided by January 30)
    • Be co-authors of the meta-analysis journal manuscript
    • Organizers of each challenge involve Challenge chairs for the author invitation to facilitate “Letter for VISA” to challenge participants
• March 15 – Camera-ready Submissions Deadline
  • (1) week for final review/acceptance by the organizers of each challenge
  • Organizers of each challenge do their ranking
• March 20 – Notification of oral versus poster presenters
  • Note that Oral papers are top-performers & interesting methods
• April 8-11 Announcement of Winners

Abstract: The Foundation Model Challenge for Ultrasound Image Analysis aims to advance the state of biomedical image analysis by fostering the development of robust, generalizable AI models for ultrasound segmentation. From a biomedical perspective, ultrasound is an indispensable diagnostic tool due to its real-time imaging, cost-effectiveness, safety, and portability—especially crucial for fetal monitoring and maternal health in low-resource settings. However, its clinical interpretation is often hindered by high inter-observer variability, low image quality, and operator dependence.

From a technical standpoint, the challenge focuses on developing foundation models that can generalize across diverse ultrasound imaging tasks and anatomical regions. Participants are tasked with creating algorithms that can accurately and reproducibly segment multiple fetal and maternal structures using a large-scale, multi-institutional dataset of over 50,000 annotated ultrasound images.

The envisioned biomedical impact is to improve diagnostic consistency, enable automated decision support, and expand access to expert-level ultrasound analysis worldwide. Technically, the challenge aims to push forward methodologies in domain-specific foundation models, standardized evaluation using BIAS-aligned metrics, and reproducible benchmarking—aligning closely with the IEEE SPS Challenge Program’s goals of supporting impactful, transparent, and community-driven signal processing research.

Abstract: Chest X-rays are one of the most widely used medical imaging modalities, yet existing AI benchmarks often overlook the challenges posed by real-world data: imbalanced disease prevalence, label noise from automated extraction, and distributional shifts across institutions.

CXR-LT 2026 is the third edition of a multi-institutional challenge designed to benchmark scalable and generalizable methods for long-tailed multi-label chest X-ray classification. It features two tasks: (1) robust prediction on a small, expert-annotated and multi-center test set and (2) open-world generalization to disease findings unseen during training.

With over 300,000 X-rays from MIMIC and MIDRC, CXR-LT 2026 provides a consistent platform for evaluating algorithmic progress in handling clinical data imbalance, noisy supervision, and cross-center generalization. The challenge aims to support the development of deployable, reproducible AI systems in medical imaging and aligns closely with the IEEE ISBI missions.”

Abstract: The proposed challenge seeks to propel the development of robust and high-quality reconstruction algorithms for magnetic particle imaging (MPI), a cutting-edge non-invasive imaging technique that visualizes superparamagnetic iron oxide nanoparticles in real time. From a biomedical perspective, MPI holds transformative potential for oncology, cardiovascular imaging, and stem cell tracking, offering detailed insights into disease processes and therapeutic monitoring. Technologically, MPI reconstruction is highly challenging due to its ill-conditioned nature, especially under low tracer concentrations, where diminished signal-to-noise ratios result in artifacts, reduced resolution, and inaccurate quantification. This challenge specifically targets field free line MPI (FFL-MPI) setups, which are gaining traction for human-scale applications and clinical translation owing to their potential for high sensitivity. The envisioned impact includes advancing technical innovations in algorithm design to enhance image quality and reliability, thereby facilitating broader biomedical adoption and improving diagnostic and therapeutic outcomes in clinical settings.

Abstract: Stroke is the second leading cause of death and the leading cause of long-term disability worldwide, imposing a tremendous burden on healthcare systems and families. As the predominant subtype, ischemic stroke accounts for nearly 71% of all stroke cases globally. Cervical artery atherosclerosis is a major etiological factor for ischemic stroke, and accurate plaque assessment has become a cornerstone of stroke prevention. Ultrasound as a real-time, low-cost, and widely accessible imaging modality, plays a key role in evaluating carotid plaques. Early detection of high-risk plaques allows for timely, targeted treatment strategies and has the potential to significantly reduce stroke incidence.

From a biological perspective, carotid plaques contribute to stroke through two primary mechanisms: progressive plaque growth causing luminal stenosis, and plaque rupture leading to thromboembolism. Both pathways result in reduced cerebral perfusion and ultimately ischemic stroke. Plaque volume and vulnerability are therefore directly associated with individual stroke risk. Accurate evaluation of plaque morphology and vulnerability, followed by appropriate treatment, is critical for effective stroke prevention and patient management.

From a technical standpoint, manual plaque morphology measurement is time-consuming and subject to considerable inter-observer variability, leading to potential bias. In addition, plaque vulnerability is influenced by multiple complex factors, making accurate diagnosis particularly challenging for junior radiologists. To address these challenges, we propose the Carotid Plaque Segmentation and Vulnerability Assessment Challenge, which aims to: (1) Segment plaque and vascular structures to quantify luminal stenosis using semi-supervised methods, thereby reducing the reliance on extensive manual annotations and mitigating inter-observer variability caused by subjective labeling differences. (2) Classify plaque vulnerability to estimate rupture risk, providing decision support for clinicians and assisting in more informed clinical decision-making.

To support this effort, we have curated a large-scale, multi-center prospective dataset comprising 1500 paired transverse–longitudinal images, collected from seven hospitals using four major ultrasound machine types. All data were acquired following standardized imaging protocols and underwent detailed segmentation and vulnerability annotations. To the best of our knowledge, this represents the first and largest paired carotid plaque ultrasound dataset with both segmentation masks and vulnerability labels, addressing the lack of publicly available paired carotid ultrasound data.

Abstract: This challenge aims to advance automated grading of ulcerative colitis (UC) from colonoscopy videos using deep learning, focusing on the Mayo Endoscopic Score (MES). It addresses the critical need for objective, reproducible, and accurate UC assessment and informative image-based description, which is currently limited by clinician subjectivity and lack of diverse datasets.

By developing and benchmarking multi-modal deep learning models developed and assessed on a multi-centre, high-quality dataset, the challenge promotes robust, generalisable AI solutions for medical imaging, in particular endoscopy and surgery —perfectly aligning with the IEEE SPS Challenge Program’s mission to foster innovative signal processing techniques that solve real-world problems in healthcare. The challenge addresses issues in two folds – 1) providing a descriptive assessment of ulcerative colitis, a known health issue that is very subjective affecting clinical diagnosis and risk to cancer; and 2) assessing novel AI frameworks rigorously that will facilitate clinical adoption.

Abstract: The primary purpose of this Challenge is to address critical bottlenecks in prenatal Congenital Heart Disease (CHD) screening via fetal echocardiography (FE). These bottlenecks include heavy reliance on highly specialized clinicians, inefficient manual annotation, diagnostic variability, data scarcity, and domain shift—all of which hinder AI-driven automation. The Challenge focuses on two core tasks: fetal cardiac ultrasound view segmentation and biometric measurement. Its goal is to drive the development of robust, clinically translatable AI algorithms. To achieve this, it prioritizes evaluating three key algorithm capabilities: segmentation/measurement accuracy, few-shot learning adaptability, and cross-clinical-setting generalization.

Abstract: Cervical cancer remains a major cause of mortality worldwide, particularly in low- and middle-resource regions where conventional Pap smears are the primary screening tool. Manual interpretation is labor-intensive and variable across experts, underscoring the need for robust, automated solutions. While AI has shown promise in cytology, progress has been limited by the absence of large, publicly available datasets of conventional smears.

The RIVA Cervical Cytology Challenge introduces the first large-scale, high-resolution, multi-expert annotated dataset of conventional Pap smears. It comprises 959 fields of view (1024 × 1024 px, 40× magnification) from 115 patients and 26,158 nucleus-level annotations across the Bethesda categories, each labeled by up to four independent cytopathologists.

The challenge features two tasks: (1) nuclei detection in complex smear backgrounds, and (2) multi-class classification of cells with explicit evaluation of uncertainty and agreement with expert variability. By establishing a reproducible benchmark and public leaderboard, RIVA will drive innovation in medical image analysis, promote trustworthy AI for clinical cytology, and ultimately support earlier and more reliable detection of precancerous and cancerous lesions.

Abstract: Leukaemia is a serious global health challenge, and its diagnosis traditionally relies on resource-intensive methods. Automating white blood cell (WBC) classification presents a promising alternative, although current models still require further validation and optimisation. This challenge benchmarks automated WBC classification on single-site microscopic blood smear images under severe class imbalance and fine-grained morphological. Acquisition and staining are standardized on one scanner; therefore, we introduce artificial noise and blur to replicate domain shift issues that may occur when different scanners or settings are used. To prevent information leakage, the held-out test set enforces patient-level separation, and train/validation use group-stratified splits to preserve minority-class coverage. The primary metric is macro-averaged F1, complemented by per-class F1, accuracy, precision, and recall. An open evaluator with a fixed submission schema ensures reproducibility. The objectives are to (i) establish a comparable benchmark, (ii) surface methods that improve rare-class reliability, and (iii) consolidate best practices via baselines and a public leader board.