What GPT-5 Means for Radiology?

RadAI Slice Newsletter

Weekly Updates in Radiology AI

Good morning, there. OpenAI's new GPT-5 model has an error rate of just 1.6% on challenging, text-based medical cases.

This new level of reliability shows that advanced AI is becoming mature enough to handle complex medical reasoning in text. For radiology, this could soon translate to powerful assistants for drafting reports and summarizing clinical data, even while its ability to interpret images remains a future goal.

Here’s what you need to know about Radiology AI last week:

• GPT-5 Arrives: A Radiologist's Guide to the 'PhD-Level' AI

• PET/MR AI detects glioblastoma recurrence earlier for oncology teams

• Carebot LLM drafts reports for musculoskeletal X-rays

• AI chest X-rays boost cost-effective osteoporosis screening

• Plus: 6 FDA approved devices & 4 new papers.

🤖 GPT-5 Arrives: A Radiologist's Guide to the 'PhD-Level' AI

Image from: OpenAI

RadAI Slice: OpenAI's new 'PhD-level' AI, GPT-5, marks a major strategic shift towards healthcare, but its immediate impact is on text, not images.

The details:

• Jumps from 0% (GPT-4o) to 46.2% on HealthBench Hard

• Error rate drops to 1.6% on hard cases (vs 15.8%)

• Key medical benchmark (HealthBench) was text-only

• Shows high scores (up to 84.2%) on general visual tests (MMMU)

• Potential uses: Reporting assistance, EMR summary

Key takeaway: GPT-5's immediate potential is augmenting the text-based parts of radiology (reporting, summaries), while direct image diagnosis remains a long-term goal requiring extensive validation.

🧠 PET/MR AI detects glioblastoma recurrence earlier for oncology teams

Image from: Health Imaging

RadAI Slice: A University of Virginia AI method outperformed traditional approaches in monitoring post-treatment glioblastoma patients.

The details:

• Analyzes multi-modal PET/MR to parse tumor vs. treatment effects

• Enables earlier intervention if glioblastoma recurs

• Current approaches require months of waiting

• Developed and tested by UVA research team

Key takeaway: AI-powered imaging offers real-time clarity for challenging high-grade brain tumors, potentially improving survival and quality of care by speeding up clinical response.

💡 Carebot LLM drafts reports for musculoskeletal X-rays

Image from: Health Imaging

RadAI Slice: A new AI assistant automates MSK X-ray reporting, reducing radiologist clerical burden.

The details:

• Focuses on MSK lesions and localization

• Reports integrate into hospital systems

• Aims for efficiency and standardization

• Plans expansion to CT, MRI, and MDR certification

Key takeaway: Radiology-specific LLMs combine image and text reasoning, setting the stage for more consistent, streamlined reporting—key as workloads rise.

📉 AI chest X-rays boost cost-effective osteoporosis screening

RadAI Slice: New evidence shows AI can screen for osteoporosis from routine chest X-rays at a lower societal cost.

The details:

• Five million simulations support US cost-effectiveness

• AI approach gained more QALYs and prevented fractures

• Treatment costs rose but overall value was high

• Study supports wider population screening

Key takeaway: AI reusing existing X-rays may close care gaps and deliver broad public health benefits with favorable economics—if health systems act.

🏛️ FDA Clearances

• K251364 - Sleepiz One+ wins FDA clearance for AI-powered, contactless home sleep and vital sign monitoring via radar sensors.

• K243473 - PRORAD ATLAS ULTRAPORTABLE is a new portable digital X-ray for extremities—ideal for remote and point-of-care imaging.

• K250356 - MoMe ARC® Wireless system receives clearance for cloud-connected ambulatory ECG arrhythmia detection with AI review.

• K251523 - Siemens Cios Spin updates enable 3D imaging and AR navigation for precise fluoroscopically guided surgery in the OR.

• K243919 - Voxel Dosimetry gets cleared: AI-powered, segmentation-based nuclear medicine tool for precise voxel-level radiation dosimetry.

• K251234 - Abbott EnSite X software maps 3D heart electrical activity for clinicians performing electrophysiological procedures.

📄 Fresh Papers

• doi:10.1016/j.media.2025.103749 - MedCLIP-SAMv2 fuses text prompts with SAM for universal, interactive medical image segmentation across modalities.

• doi:10.1016/j.media.2025.103748 - Recurrent Inference Image Registration (RIIR) framework raises accuracy and data efficiency for multi-modal image registration.

• doi:10.1016/j.compmedimag.2025.102617 - Multimodal ensemble learning improves early atherosclerosis assessment by integrating clinical risk and ultrasound data.

• doi:10.1007/s00330-025-11920-7 - DL-based image enhancement boosts MRI quality and sensitivity for small brain metastases, especially under 5 mm.

📰 Everything else in Radiology AI last week

• Regulatory and ethical shifts: Illinois bans AI in therapy, Mayo runs 100+ clinical algorithms, AMA weighs in.

• Federal policy shifts, GenAI commercialization models, and gaps in AI governance reshape radiology AI's direction.

• UnitedHealthcare's nationwide Heartflow AI coverage cements CT-based plaque analysis as reimbursed clinical standard.

• Harvard AI predicts radiology staffing needs a day ahead, helping to ease burnout and allocate resources.

• Open-source Sybil AI excels in predicting lung cancer risk in Asian heavy smokers using low-dose CT scans.

That's it for today!

👋 We’d love your thoughts. Just hit Reply and tell us what you want to see in the next Slice!

See you next week.