🚀We are excited to release the largest Pan-Cancer CT Segmentation Dataset (17000+ labeled cases) and launch the Foundation Models for Pan-Cancer Segmentation in CT Images Challenge! 📦 The dataset brings together publicly available CT cancer datasets with lesion annotations: huggingface.co/datasets/FLAR… 🏆 Challenge and submission: codabench.org/competitions/7… Help us build more scalable and generalizable AI models for cancer imaging and please share this challenge with anyone who may be interested! @MICCAI_Society @bias_sig @BoWang87 @bhaibeka @Zongwei_Zhou @yuyinzhou_cs

BodyMaps AI is now live. We're recruiting volunteers across radiology, radiation oncology, and surgical departments for a reader study on how AI assists early cancer detection, tumor contouring, and surgery/radiotherapy planning. Try it here: bodymaps.wse.jhu.edu/ — DM me to join. Volunteers would earn authorship in our clinical papers. @JohnsHopkins @HopkinsMedicine

Introducing JEV-27B-VL. We believe it is the world’s first open-weight, near-SOTA multimodal decision model — built with AutoTrust AI’s Blocks of Experts (BoE) recipe. It does not just describe what it sees. It turns visual state into calibrated action probabilities, then selects what to do next. In this demo, it sees Mario’s position, movement, obstacles, and timing — then chooses whether to run, jump, or do both. See → decide → act.

MICCAI 2026 papers are now available through open access! 🎉 papers.miccai.org/miccai-202… Ahead of the conference, I (and GPT-6) analyzed 3,048 MICCAI paper titles from 2024–2026 and put the results into the GIF below. Some trends stood out 👇 🧠 Reasoning and agents are gaining visibility fast. Titles mentioning reasoning grew from 2 → 9 → 34. Agents/agentic terminology rose from 1 → 8 → 30. Both started from small bases, but the increases are striking. 🌊 Diffusion is losing title share as flow matching emerges. Diffusion mentions fell from 88 to 68, while their share dropped from 10.3% to 5.8%. Explicit flow-matching mentions went from 0 → 4 → 18. An interesting parallel (not evidence that one is replacing the other). 🏗️ Foundation-model terminology expanded, then leveled off. “Foundation” appeared in 22 → 40 → 45 titles. Its share stayed near 3.9% in 2025 and 2026. 📉 Architecture names have different trajectories. Transformer mentions declined from 35 → 29 → 26, while Mamba peaked in 2025: 12 → 23 → 20. Less visibility in titles does not necessarily mean less use inside the papers. 🎯 Segmentation remains a major focus. Segmentation papers increased from 198 to 232, even as their share of the growing conference fell from 23.1% to 19.9%. A good reminder that rising counts and falling shares can happen together. 🧩 Multimodal research remains prominent. Titles matching multimodal terminology increased from 47 in 2024 to 80 in 2025, holding at 79 in 2026. I also maintain MICCAI-OpenSourcePapers, a collection of MICCAI papers with code: github.com/JunMa11/MICCAI-Op… Thanks @MICCAI_Society for making the papers open access. #MICCAI2026 #MICCAI #MedicalAI #MedicalImaging #OpenScience

Curating clinical variables from free-text notes is tedious. General LLMs can help, but processing thousands of notes can be slow and costly. Inspired by Jev and the open-source community, we’re releasing MedJev to turn clinical notes into structured fields on consumer GPUs. A 0.8B model + 43 MB LoRA adapter, trained to extract 11 predefined clinical variables. On our benchmark of 2,895 held-out notes: ⚡ 62 ms per note, 4.6× faster than the Jev API 🎯 87.8% accuracy vs. 60.2% for Jev 🏠 Runs locally, patient text stays on your network Data & Code: github.com/JunMa11/MedJev Model: huggingface.co/junma/MedJev-…

@ScienceMagazinea just published a fascinating AI model that can detect 146 diseases from abdominal CT scans. I built a demo so you can try this impressive model yourself 👇 huggingface.co/spaces/junma/… paper: science.org/doi/10.1126/scie…

UHN was just named by @newsweek as the no. 1 hospital in Canada for: 🥇 Cardiology 🥇 Cardiac Surgery 🥇 Endocrinology 🥇 Gastroenterology 🥇 Nephrology 🥇 Pulmonology 🥇 Oncology 🥇 Neurosurgery Learn more about the 2027 World's Best Specialized Hospitals: rankings.newsweek.com/worlds…

A brief summary of the nnFamily: toward out-of-the-box solutions for medical image analysis nnU-Net showed that strong medical image analysis does not always require inventing a new architecture. Instead, it systematically adapts preprocessing, network configuration, training, inference, and post-processing to each dataset, making high-quality biomedical image segmentation accessible with minimal manual intervention. With 10,000+ citations, nnU-Net has become one of the most influential and widely used frameworks in medical imaging AI. Its success inspired a broader idea: Can the same self-configuring, robust, and user-friendly philosophy be extended beyond semantic segmentation? A growing “nnFamily” is now bringing this vision to many areas of medical image analysis: 🔹 nnU-Net — Semantic segmentation @FabianIsensee1 @maierhein The foundation of the family: a self-configuring framework that automatically adapts a complete segmentation pipeline to new biomedical datasets. nature.com/articles/s41592-0… 🔹 nnDetection — Medical object detection Extends the self-configuring philosophy to the localization and classification of objects in medical images, automatically adapting the detection pipeline to different datasets. arxiv.org/abs/2106.00817 🔹 nnInteractive — Interactive 3D segmentation A promptable, open-set framework for volumetric medical images. It supports points, scribbles, bounding boxes, and lasso prompts to efficiently generate full 3D segmentations across diverse structures and modalities. arxiv.org/abs/2503.08373 🔹 nnLandmark — 3D landmark detection A self-configuring framework for detecting anatomical landmarks in 3D medical images, designed to deliver strong performance on new datasets without task-specific hyperparameter tuning. arxiv.org/abs/2504.06742 🔹 nnActive — Active learning A standardized framework for studying and applying active learning in 3D biomedical segmentation. It supports partial annotations and patch-based sample selection, helping researchers investigate how annotation effort can be reduced effectively. arxiv.org/abs/2511.19183 🔹 nnMIL — Computational pathology @xiangzLuo A generalizable multiple-instance learning framework that connects patch-level pathology foundation models with robust slide-level clinical predictions, including diagnosis, subtyping, biomarker detection, and prognosis. arxiv.org/abs/2511.14907 🔹 nnsyn — Medical image synthesis A self-configured and extensible framework for cross-modality medical image synthesis, providing streamlined preprocessing, training, and inference while supporting masks, perceptual supervision, and multimodal inputs. github.com/aehrc/nnsyn 🔹 nnClass — Medical image classification Extends the nnU-Net ecosystem toward binary and multiclass classification, with support for multimodal CT, PET, and MRI data, stratified cross-validation, sliding-window inference, test-time augmentation, and model ensembling. github.com/bowang-lab/nnClas… Together, these projects point toward a broader future for medical imaging AI: ✅ Less manual pipeline engineering ✅ Strong, reproducible baselines ✅ Automatic adaptation to new datasets ✅ Standardized training and evaluation ✅ Support across segmentation, detection, classification, interaction, landmark localization, synthesis, pathology, and active learning ✅ More accessible AI tools for researchers and clinical domain experts What should the next member of the nnFamily address? #MedicalAI #MedicalImaging #nnUNet #DeepLearning #ComputerVision #AutoML #RadiologyAI #ComputationalPathology

From task-specific U-Nets to universal promptable models: glad to share our perspective on where the field has been and where it’s going!

Just noticed two of our CVPR 2025 papers are listed in PaperDigest’s Top-15 Most Influential CVPR 2025 Papers: #11 VA-VAE / LightningDiT github.com/hustvl/lightningd… #14 DiffusionDrive github.com/hustvl/diffusiond… Grateful to all coauthors and collaborators. paperdigest.org/2026/03/most…

#CVPR2026 Medical Vision Highlights 6: "OralGPT-Plus: Learning to Use Visual Tools via Reinforcement Learning for Panoramic X-ray Analysis" paper: openaccess.thecvf.com/conten… code: github.com/isjinghao/OralGPT The question: panoramic dental X-rays need fine spatial reasoning, bilateral-symmetry checks, and multi-step verification, but single-pass VLMs answer in one shot and can’t revisit an ambiguous region. Can a VLM instead act like a dentist who zooms and compares? Key idea: Design, an agentic thought, action, observation loop with two visual tools: “Zoom-In” for high-res inspection of a suspicious area, and “Mirror-In”, which retrieves the contralateral region across the midline so the model can compare symmetric teeth. DentalProbe: a 5k-image dataset with 8k+ expert-guided diagnostic trajectories (global scan, proposal, local exam, symmetry check), built via a multi-agent refinement pipeline with dentist review. Highlights: on their MMOral-X benchmark (300 open-ended Q&A, 686 boxes, 3 difficulty levels), OralGPT-Plus-7B scores 45.35 overall vs GPT-5’s 42.34, also leading on MMOral-OPG. Ablations: removing “Mirror-In” drops Simple accuracy 43 to 34.7; the conditioned reward is what prevents reward-hacking (tool-spam).

#CVPR2026 Medical Vision Highlights 5: "Medic-AD: Towards Medical Vision-Language Model’s Clinical Intelligence" The question: medical VLMs know a lot, but real workflows need three specific things, accurate lesion detection, reliable tracking of change over time, and transparent visual evidence. How to build those in rather than hope they emerge from generic instruction tuning? Key Idea: Design, a stage-wise framework on a Lingshu-7B backbone that mirrors a clinical workflow: detect, compare, explain. Stage 1: learnable anomaly-aware tokens (abnormal vs normal system tokens via cross-attention) push the model toward lesion regions. Stage 2: inter-image tokens encode change between prior and current scans, worsened / improved / unchanged. Stage 3: an explainability head, tokens fused with visual features through a ConvNeXt decoder produce heatmaps grounded on the lesions driving the prediction, rather than a generic post-hoc saliency map. Highlights: zero-shot anomaly detection avg F1 91.6 (Lingshu 88.7, GPT-4o 69.4) across 4 unseen modalities; MMXU symptom-tracking overall 0.655 (Lingshu 0.620); validated on a real 300-patient longitudinal hospital set.

#CVPR2026 Medical Vision Highlights 4: "Efficient Unrolled Networks for Large-Scale 3D Inverse Problems" paper: arxiv.org/abs/2601.02141 code: romainvo.github.io/efficient… The question: unrolled networks that embed the imaging operator A into the architecture give SOTA reconstruction, but training them needs the full-resolution forward pass. In 3D (e.g., 501³ voxels) that blows past a single GPU’s memory. How to scale them? Key idea 1: domain partitioning. Decompose the volume into small patches and solve for a patch given the rest as known context, even when A couples all voxels. This extends patch-based training to operator-aware unrolling. Key idea 2: normal-operator approximation. Replace AᵀA with a product of diagonal and circulant matrices, evaluated via FFT. Neatly, the factorization parameters can be fit by gradient descent on random vectors, no problem-specific training data needed. Highlights: on Walnut CBCT (501³) and Calgary-Campinas multi-coil MRI, it trains and runs on a single GPU where the standard 3D unrolled network simply OOMs. It reaches the best reconstruction among compared methods, with ~30% less training time when the operator approximation is added.

Had a wonderful time at @CVPR 2026, so many thoughtful conversations and inspiring work. Here are six medical vision papers I'd like to share, one per core task: 1. Segmentation: VoxTell (DKFZ). Universal 3D segmentation driven by free text, from a single word to a full clinical sentence. Trained on 62K+ CT/MRI/PET volumes; the key idea is fusing text into the decoder at every scale, not just at the end. 2. Classification: AnyMC3D (Siemens Healthineers). Instead of new 3D foundation models, it adapts frozen 2D ones with ~1M params/task. Careful study of three recurring pitfalls in the field: data-regime bias, weak adaptation, and too-narrow task coverage. 3. Registration: SGDIR (U Alberta). Diffeomorphic registration using a single semigroup regularizer, with a proof that it makes the network learn an ODE flow, so invertibility and cycle consistency come for free, no scaling-and-squaring. 4. Reconstruction: Efficient Unrolled Networks (CNRS/ENS Lyon). Two ideas (domain partitioning plus a normal-operator approximation via FFT) let you train operator-aware unrolled networks for 3D CBCT and multi-coil MRI on a single GPU. 5. Medical VLM: Medic-AD (SNU/Samsung/NVIDIA). A stage-wise design that turns broad VLM knowledge into three concrete clinical skills: anomaly detection, longitudinal symptom tracking, and grounded heatmap explanations. 6. Agent: OralGPT-Plus (HKUST-GZ/HKU/PKU). An agentic VLM for panoramic dental X-rays that learns to "zoom in" and "mirror in" (compare symmetric teeth), trained with reinspection-driven RL. A short series of notes follows. Corrections welcome, I'm sharing to learn, not to rank😀

This signal is hard to miss for those of us working at the intersection of AI and healthcare! Health and life sciences are named the #1 priority sector, and the country's very first AI Mission is in healthcare. A few things I'm especially glad to see: - Public supercomputer for researchers and SMEs - A genuine commitment to open-source AI, the foundation of reproducible, adaptable medical AI - Investment in retaining talent, because we train world-class people and too often watch them build elsewhere

Prof. Rajpurkar @pranavrajpurkar is presenting Routing to Autonomous AI for Medicine! Welcome to join us at room 607. @CVPR @BoWang87 @yuyinzhou_cs