Ontology-Based Concept Distillation: Interpretable Retrieval for Radiology Reports скачать в хорошем качестве

Ontology-Based Concept Distillation: Interpretable Retrieval for Radiology Reports

Ontology-Based Concept Distillation: Interpretable Retrieval for Radiology Reports YouTube Description (with unicode formatting): Can ontology-driven AI outperform black-box embeddings in medical imaging? 🩻✨ In this video, we explore Ontology-Based Concept Distillation, a novel framework that leverages UMLS concepts for radiology report retrieval and disease labeling. Developed by researchers at Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) and Imperial College London, this approach is pushing the boundaries of explainable Medical AI. Key Highlights: 💡 Entity Extraction: Combining RadGraph-XL and SapBERT to capture standardized medical concepts. ⚖️ Smarter Similarity: Modified Tversky Index handling synonymy, negation, and hierarchies. 📊 Improved Retrieval: Outperforming CLIP, Bio-ClinicalBERT, and CXR-BERT in long-tail classification on MIMIC-CXR. 🏷️ Ontology-backed Labels: More accurate than CheXpert and modern LLMs for disease labeling. Takeaways: 🔬 Ontology-based methods deliver higher AUROC than state-of-the-art embeddings. 🧠 Provides interpretable, task-adaptive similarity for clinical AI. 🌍 Opens new pathways for safer, scalable, and transparent healthcare AI systems. 📄 Full paper: https://arxiv.org/abs/2508.19915 Subscribe for more deep dives into the latest breakthroughs in Medical AI! 🚀 #MedicalAI #Radiology #ExplainableAI #AIResearch #HealthTech