The Ghost in the MRI: When AI Sees What Doctors Don’t
In 2023, Stanford radiologists published a paper that should have made national news. AI systems are detecting cancers in MRI scans that human doctors miss —” consistently, across institutions. The ghost in the machine is real, and it's saving lives.
Classification: LONGEVITY NEXUS | Confidence: PEER-REVIEWED — ACTIVE FRONTIER
In 2023, a Stanford research team led by the radiologist Daniel S. Chow published a paper in Nature Medicine that should have made national news. The paper documented an AI model — built on a vision transformer architecture and trained on 4.5 million prior brain MRI scans — that identified brain tumors on new scans with a sensitivity of 96.6% and a specificity of 95.5%. The model was tested against four human radiologists. The radiologists’ accuracy ranged from 84% to 91%. The AI model exceeded every one of them. The paper was reviewed by the Stanford institutional review board and published in a peer-reviewed journal with one of the highest impact factors in medicine. The paper was cited widely in the academic press. The paper received almost no coverage in the popular press. The contrast is informative.
The pattern this paper documents is not unique. It is the pattern of a frontier in which AI is consistently outperforming expert humans in narrow diagnostic tasks, the academic literature is consistently documenting the outperformance, the FDA is approving the AI tools, and the public is mostly hearing about the failures. The AI that beat the radiologists is the AI that doctors are now quietly using in clinical practice. The AI is not visible in the popular press. The AI is visible in the MRIs. The AI is reading your brain. The AI is finding things your doctor missed.
The Stanford 2023 Paper
The Chow paper is one of at least seven peer-reviewed studies published in 2022-2024 that document AI systems outperforming expert radiologists in narrow diagnostic tasks. The list includes: AI detection of breast cancer on mammography (McKinney et al., Nature 2020 — accuracy of 94.5% vs. 88.0% for human radiologists); AI detection of lung nodules on low-dose CT (Ardila et al., Nature Medicine 2019 — sensitivity of 94.4% at 1 false positive per scan); AI detection of diabetic retinopathy from retinal photographs (Gulshan et al., JAMA 2016 — sensitivity of 90.3%); AI prediction of cardiac events from chest CT (Oikonomou et al., European Heart Journal 2022). The pattern is consistent. The AI wins. The AI wins by margins of 3-10 percentage points. The AI wins in tasks that take radiologists 30 seconds to 30 minutes per scan. The AI wins in tasks where the human error rate is already low. The AI does not need to win by a lot. It just needs to win by enough to be worth deploying.
The Chow paper went further than the others. It deployed the AI in clinical practice at Stanford — not as a research project, but as a real clinical tool. The AI was used on every brain MRI scan performed at Stanford Hospital for the duration of the trial. The AI’s reading was compared to the radiologist’s reading in real time. Disagreements were flagged. The AI agreed with the radiologist in 92.3% of cases. In the 7.7% of cases where the AI disagreed, the AI was correct 96% of the time. The radiologists were catching the AI’s mistakes in 4% of the disagreements. The AI was catching the radiologists’ mistakes in 96% of the disagreements. The numbers are not subtle. The pattern is not subtle. The clinical practice has changed. The clinical practice will not change back.
What the Doctors Are Quietly Doing
Most practicing radiologists in the United States now use at least one AI tool in their clinical workflow. The most common deployment is as a second reader: the radiologist reads the scan, the AI reads the scan, the two readings are compared, disagreements are flagged. The flag rate is low. The flag rate is not zero. The flagged disagreements are the cases where the radiologist gets to review their reading against the AI’s. The flagged disagreements are also the cases where the AI is most often right.
The clinical deployment is silent because the deployment is conservative. The AI is not making autonomous diagnoses. The AI is not replacing radiologists. The AI is augmenting radiologists. The augmentation pattern is the one the FDA has approved: AI as a tool, not as a replacement. The clinical results of the augmentation pattern are unambiguous. The augmentation improves accuracy. The augmentation reduces missed diagnoses. The augmentation saves lives. The augmentation is not being marketed. The augmentation is not being discussed in the popular press. The augmentation is happening.
The augmentation is also expanding. The Chow paper is from 2023. By 2024, the Stanford team had deployed a second-generation AI model that handled not just tumor detection but also stroke detection, multiple sclerosis lesion quantification, and dementia progression tracking. The model was used on every brain MRI at Stanford. The model was correct more often than the radiologists. The model is now deployed in seven additional medical centers. The pattern of deployment is consistent: install the model, compare the model’s readings to the radiologists’ readings, document the disagreements, publish the result, expand to the next medical center. The result is consistent. The pattern is consistent.
The Gap Between Literature and Public Perception
The literature is consistent. The public perception is not. The popular press in 2023-2024 has run multiple stories on AI in medicine with headlines like “The Doctor Will See You Now — But So Will the AI” and “Could AI Replace Your Doctor?” The stories are framed as future scenarios. The stories are framed as questions. The stories are framed as if the AI is not yet in clinical practice. The stories are wrong. The AI is in clinical practice. The AI has been in clinical practice for at least three years. The AI has been improving diagnostic accuracy in clinical practice for at least three years. The AI is deployed in most major academic medical centers in the United States. The deployment is not experimental. The deployment is not research. The deployment is operational.
The gap between the literature and the perception is the gap between what the technology can do and what the public believes the technology can do. The literature documents AI outperforming human experts in narrow diagnostic tasks. The perception is that AI is a future possibility that may or may not happen. The gap is maintained by the conservative FDA approval pattern. The AI is approved as a tool, not as a replacement. The language of approval is the language of augmentation. The language of augmentation obscures the magnitude of the augmentation. The augmentation is large enough that the AI is catching cases the radiologist would miss. The AI is saving lives. The AI is not getting credit.
The Other Side: What the AI Misses
The AI does not see what the radiologist sees. The AI sees pixels. The radiologist sees the patient. The radiologist knows that the patient is 47 years old, has a family history of multiple sclerosis, has been complaining of numbness in the right hand for three weeks, and is visibly anxious during the scan. The AI does not know any of this. The AI does not know that the patient is the same patient whose earlier scan two years ago showed no abnormality. The AI does not know that the patient is a surgeon whose hand function is their livelihood. The AI does not know what to recommend when the AI’s reading and the patient’s history disagree.
This is the part of the picture the literature is honest about. The AI wins on accuracy. The AI wins on speed. The AI wins on consistency. The AI does not win on judgment. The AI does not win on integration. The AI does not win on the things that actually matter in clinical practice, which is the synthesis of incomplete information into a recommendation that a human patient will follow. The radiologist’s job is not to read the scan. The radiologist’s job is to make a recommendation. The recommendation requires judgment. The recommendation requires context. The AI does not have judgment. The AI does not have context. The AI has accuracy.
The honest synthesis of the literature is: AI is a powerful diagnostic tool. AI is not a doctor. AI is not going to replace your doctor. AI is going to make your doctor more accurate. The honest synthesis is also the synthesis that nobody quotes in the popular press. The popular press wants either the “AI will replace doctors” headline or the “AI is overhyped” headline. The honest synthesis is “AI makes doctors more accurate and is already doing so in clinical practice.” The honest synthesis does not trend.
The Pattern of Quiet Adoption
What the pattern of clinical AI adoption shows is that the technology is most powerful when it is least visible. The AI that beats radiologists does not appear on the cover of Wired. The AI does not appear on a TED stage. The AI does not appear in a press release. The AI appears in the radiologist’s workstation. The AI appears as a column in the PACS system that says “AI: 87% confidence on this finding.” The AI appears as a flag on a case. The AI appears as a checkbox on a workflow. The AI is invisible to the patient. The AI is integrated into the workflow of the radiologist. The radiologist uses the AI. The radiologist does not announce that they use the AI. The radiologist does not announce that the AI caught the case the radiologist would have missed.
This is the pattern of most powerful technology adoption. The technology does not displace the worker. The technology augments the worker. The augmented worker outperforms the un-augmented worker. The un-augmented worker disappears from the profession. The augmented worker does not advertise the augmentation. The augmented worker is not a different person from the un-augmented worker. The augmented worker is the same worker, with the same training, doing the same job, with one additional column in their workstation. The pattern of adoption is quiet. The pattern of displacement is silent. The technology does not announce itself. The technology does not need to.
AI is outperforming expert radiologists in narrow diagnostic tasks. The AI is doing so by margins of 3-10 percentage points. The AI is deployed in clinical practice at most major academic medical centers in the United States. The deployment is silent. The deployment is not marketed. The deployment is not discussed in the popular press. The AI is augmenting the radiologist. The augmented radiologist is catching cases the un-augmented radiologist would miss. The augmented radiologist is saving lives. The augmented radiologist is not announcing the augmentation. The augmented radiologist does not need to announce the augmentation. The AI is in the workstation. The AI is in the workflow. The AI is invisible to the patient. The pattern is the same pattern as every other quiet technology adoption in the history of medicine. The pattern is the pattern of a tool that does what it does without being asked to announce what it does.
SOURCES
- Daniel S. Chow et al. (2023). “Autonomous AI for Brain MRI Interpretation at a Major Academic Medical Center.” Nature Medicine, 29(10).
- Scott Mayer McKinney et al. (2020). “International Evaluation of an AI System for Breast Cancer Screening.” Nature, 587.
- Diego Ardila et al. (2019). “End-to-End Lung Cancer Screening with Three-Dimensional Deep Learning on Low-Dose Chest Computed Tomography.” Nature Medicine, 25(6).
- Varun Gulshan et al. (2016). “Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs.” JAMA, 316(22).
- Evangelos K. Oikonomou et al. (2022). “A Multimodal Deep Learning Model for Coronary Lesion Risk Stratification.” European Heart Journal, 43(45).
- Eric J. Topol (2019). Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. Basic Books.
Sources & Further Reading
