Nassir Navab

Achieving anatomically consistent multi-view fusion improves stenosis reporting accuracy, overcoming the limitations of single-view models.

Achieving consistent voxel spacing can dramatically improve medical image segmentation quality, leading to better diagnostic and surgical outcomes.

Generating realistic patient embeddings can yield performance on par with full datasets, even in scenarios with missing classes.

PANY outperforms existing model-free methods by over 20% in pose accuracy, even in challenging conditions with limited reference overlap.

P-JEPA achieves state-of-the-art action classification on long procedural videos while using an order of magnitude fewer parameters than existing models.

Disentangling ego-motion from environmental dynamics allows FR3D to achieve unprecedented geometric consistency in future 3D reconstructions.

Traditional scene graph methods falter in capturing the temporal structure of OR activities, while a new vision-only model achieves superior performance in multi-role action recognition.

Self-supervised depth estimation gets a boost: SA4Depth aligns scene scales between pose and depth networks, leading to substantial improvements in depth prediction without sacrificing inference speed.

A simple "resilience" metric turns out to be surprisingly effective at identifying failure cases in NCA-based medical image segmentation, improving trust without retraining.

Forget full automation; the future of medical robotics is "Dyadic Partnership," where AI and clinicians collaborate as equals, leveraging generative AI and intuitive interfaces for shared decision-making.

Ditch anatomical segmentations: this method tracks disease progression by watching how MRI sequence vectors "flow" across a baseline energy landscape learned from a single scan.

Surgical robots can now reason about 4D spatiotemporal relationships in laparoscopic videos, without any additional training, by simply combining existing 2D MLLMs with a novel 3D computer vision pipeline.

Radiology report generation models can now verbalize calibrated confidence estimates, enabling targeted radiologist review of potentially hallucinated findings.

An RL-aligned LLM can outperform expert toxicologists in identifying ingested substances from heterogeneous clinical data, suggesting a path to AI-assisted decision-making in high-stakes medical environments.

Forget brittle, fixed robotic US scanning procedures: this LLM-powered agent dynamically interprets guidelines and adapts to real-time observations, enabling autonomous scanning across diverse anatomical targets.

Instruction-tuned LLMs can mine free-text radiology reports to create a knowledge base that significantly improves the accuracy of structured report generation, especially for rare and detailed findings.

Ditch the flat scene graphs: TopoOR models surgical environments as higher-order topological structures, unlocking superior performance in safety-critical tasks by preserving complex relationships and multimodal data.

Achieve 80% better shape completion for ultrasound reconstruction by implicitly modeling acoustic interactions, eliminating the need for anatomical labels during inference.

Achieve robust dual-arm robotic ultrasound-guided interventions by learning personalized expert strategies from limited demonstrations using a phase-aware imitation learning policy.

By mimicking how pathologists dynamically integrate global and cellular evidence across multiple magnifications, MMNavAgent significantly boosts WSI diagnostic performance.

Achieve autonomous robotic ultrasound examinations with RAG-RUSS, an interpretable framework that explains its actions and generalizes well even with limited training data.