Episode

From MRI to World Models: How AI Is Changing What We See

Podcast
AI Engineering Podcast
Published
Oct 27, 2025
Duration seconds
2931
Processing state
processed
Canonical source
https://www.aiengineeringpodcast.com/ai-vision-understanding-episode-66
Audio
https://op3.dev/e/dts.podtrac.com/redirect.mp3/serve.podhome.fm/episode/f6ff0caa-931b-4c08-bfdd-08dc7f5cd336/638971258438915274b33aa179-f83a-4214-9c01-f79ff9c5263e.mp3
JSON
/v1/public/podcasts/ai-engineering-podcast/episodes/from-mri-to-world-models-how-ai-is-changing-what-we-see
Markdown
/podcast/ai-engineering-podcast/from-mri-to-world-models-how-ai-is-changing-what-we-see.md

Actions

  • POST https://stenobird.com/v1/public/podcasts/ai-engineering-podcast/episodes/from-mri-to-world-models-how-ai-is-changing-what-we-see/transcription-requests
    Idempotently request low-priority transcript generation for this episode.
  • GET https://stenobird.com/podcast/ai-engineering-podcast/from-mri-to-world-models-how-ai-is-changing-what-we-see.md
    Read the agent-friendly Markdown representation of this episode resource.

Summary

AI is moving beyond simple image interpretation toward 'upstream' applications that redefine how we capture physical reality. By integrating physics-guided constraints into neural networks, researchers are creating more accurate, efficient, and accessible medical imaging.

Topics

  • Medical Imaging
  • Physics-Guided Neural Networks
  • World Models
  • Radiology
  • Computer Vision
  • Deep Learning
  • MRI Reconstruction
  • AI Explainability

Highlights

  • Main idea: 'Upstream' AI focuses on changing how we measure the world, rather than just interpreting existing data
  • Practical takeaway: Incorporating physical laws into neural networks ensures that AI-generated reconstructions remain physically plausible
  • Failure mode: Relying solely on scaling LLMs may fail to capture the complex spatial and physical context required for medical-grade vision
  • Main idea: The future of radiology involves AI agents performing initial screenings and flagging abnormalities for human judgment
  • Practical takeaway: Using physics-guided networks can enable faster, cheaper, and more accessible imaging modalities

Chapters

  1. 4:30 Defining the Image: A discussion on viewing images as spatially organized maps of information rather than just pixels.
  2. 8:20 The Role of Domain Expertise: How integrating human expertise and physical constraints prevents AI from generating unrealistic results.
  3. 12:45 Upstream vs. Downstream AI: Distinguishing between AI that interprets images and AI that fundamentally changes how we capture data.
  4. 20:00 Beyond Scaling LLMs: Why foundation models need to incorporate CT and MRI data to move toward true world models.
  5. 24:10 The Challenge of Explainability: Addressing the lack of interpretability in deep learning within high-risk medical environments.
  6. 27:40 The Future of Clinical Workflow: A vision of AI handling routine exclusions and human clinicians focusing on complex diagnostic decisions.
  7. 45:15 The Era of Synthetic Content: Reflecting on a future where human-created content becomes the minority in a sea of AI-generated imagery.