I am uncomfortable with the proliferation of AI-generated radiology images in social media education. They're not labeled as AI and undiscerning viewers are routinely fooled into thinking they're real.

The most accurate description of an ER visit lol

Using AI to analyze whole-body MRI scans from more than 66,000 people, researchers showed that skeletal muscle quality and fat distribution are powerful predictors of diabetes, cardiovascular events and mortality. The findings suggest advanced imaging can reveal hidden health risks that traditional measures like BMI often miss. The study also introduces an open-source AI tool that allows clinicians and researchers to extract powerful insights from scans already being performed—opening the door to earlier detection, better risk stratification and more personalized care. Read the full story: bit.ly/4umqtcK #MSK #MRI

With all due respect most radiologists are already readmaxxing images without AI; AI will likely marginally increase their volume but they’re essentially time capped. So with all due respect, who’s going to validate the AI reads? More radiologists is the answer

AI won't take over radiology... until it takes over radiology. Despite the hype or the protest, evidence is slowly mounting that that day will eventually come. The latest study in BMJ Gut is compelling. A Mayo Clinic team built REDMOD, an AI framework designed to detect pancreatic ductal adenocarcinoma at stage 0 — before any visible mass appears on CT. Before any radiologist, no matter how experienced, could possibly identify it. They tested it on 1,462 patients in a multi-institutional cohort designed to simulate real-world low-prevalence screening (6:1 control-to-case ratio). The results: AI detected visually occult pancreatic cancer with 73% sensitivity. The radiologists achieved 38.9%. That's nearly double. At lead times beyond 24 months — meaning more than two years before clinical diagnosis — the gap widened to nearly threefold: 68% vs 23%. Overall AUC: 0.82 for the AI vs 0.69 for the radiologists (p<0.001). This isn't an AI reading the same images faster. This is an AI seeing what human eyes structurally cannot. The cancers in this study were confirmed as imaging-occult prospectively read as negative by board-certified radiologists, then independently re-reviewed and confirmed negative again. There was nothing to see. REDMOD found the signal anyway, embedded in subvisual textural patterns across the pancreatic parenchyma. And the signal was stable. Test-retest concordance hit 90-92% across serial scans, with one patient correctly flagged 1.8 years before diagnosis and tracked through an evolving radiomic signature that preceded any visible tumor. The interobserver agreement between the two radiologists? Kappa of 0.22. The AI was not only more accurate, it was more consistent than the humans by an order of magnitude. Pancreatic cancer kills 85% of patients at current detection timelines. Modeling studies suggest that shifting even a fraction of diagnoses from late to localized stage would more than double survival rates. REDMOD offers a median 475-day detection window. That is the difference between palliative care and curative surgery. The profession's instinct is to frame AI as a tool that augments the radiologist. And for most imaging tasks today, that framing holds. But studies like this reveal a different trajectory, one where the most consequential diagnostic capability isn't augmentation. It's perception beyond the human visual threshold. Radiology will look very different in well under ten years. The profession can lead that transformation or get restructured by it.

gpt-5.5 prompt for codex seems to have a duplicated line trying to get it to not talk about creatures? Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query. [...] Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query gh link: github.com/openai/codex/blob…

We have numbers. Big bump to plain films at the expense of body CT and neuro CT/MRI as well as a general devaluing of ED work. On the breast side, a huge cut to screeners for a smaller increase for diagnostics. For IR, no delta. GI fluoro still isn’t fair. benwhite.com/radiology/rp-tb…

New guidance from the American College of Physicians (ACP) says all asymptomatic, average-risk females ages 50 to 74 should receive biennial screening mammography for breast cancer. Females between the ages of 40 and 49 should discuss with their doctor their risk for breast cancer and the benefits and harms of screening. This is because harms of screening such as false positive results, psychological distress because of it, overdiagnosis, overtreatment, additional testing, and radiation exposure may outweigh the uncertain benefits in this population. Read more: bit.ly/4mBnHOh

AI isn’t coming to medicine—it’s already transforming it. Use among physicians has surged from 38% to 81%, helping cut admin burden and refocus care on patients. The priority now: keep it transparent, accountable and physician-led. spr.ly/6011B6QyEL