Medicare denied 53% of radiology AI claims since 2018

RadAI Slice Newsletter

Weekly Updates in Radiology AI

Good morning, there. Medicare denied $16.4 million—53%—of radiology AI claims from 2018 to 2023.

This national analysis shows a persistent barrier to AI reimbursement in radiology, impacting ROI and adoption. High denial rates may slow integration of AI into clinical workflows. Addressing policy clarity and coverage consistency is key for sustainable AI implementation.

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

• Medicare Denied Over Half of Radiology AI Claims

• GE HealthCare Acquires icometrix for MRI Neuro-AI Integration

• Major Study Reveals Barriers to NHS AI Chest Tool Deployment

• Stanford, Mayo: LLMs Flag Critical Findings in Radiology Reports

• Plus: 3 newly released datasets, 3 FDA approved devices & 4 new papers.

🆕 RadAI Chat Waitlist

We’re building RadAI Chat — a side-by-side case chat where you can upload once, run multiple AI models at the same time, and instantly compare results with native DICOM support.

Exclusively for subscribers. Watch the demo and join early.

💵 Medicare Denied Over Half of Radiology AI Claims

Image from: Radiology Business

RadAI Slice: A new longitudinal analysis highlights ongoing denial rates for radiology AI services billed to Medicare.

The details:

• 83,392 AI service claims submitted to Medicare 2018–2023

• 53,857 (≈53%) denied, totaling $16.4M in rejected payments

• 39,535 (≈47%) AI claims paid, for $8M reimbursement

• More than 1,200 FDA-cleared AI devices now in use

Key takeaway: High denial rates create reimbursement uncertainty, slowing radiology AI adoption. Policy clarity is urgently needed for sustainable clinical integration.

🧠 GE HealthCare Acquires icometrix for MRI Neuro-AI Integration

RadAI Slice: GE HealthCare is acquiring icometrix, a leader in MRI neurology AI, for tighter workflow and analytics integration.

The details:

• icometrix technology addresses Alzheimer’s and MS, including ARIA detection

• Planned integration with GE and vendor-agnostic MRI networks

• icometrix partnerships include Siemens, Nuance, and ACR

• Consolidation trend continues as OEMs expand AI portfolios

Key takeaway: Consolidation enables OEMs to deliver end-to-end specialty AI integrated in radiologist workflow—critical as neuroimaging volumes rise.

🏥 Major Study Reveals Barriers to NHS AI Chest Tool Deployment

RadAI Slice: A multicenter NHS study outlines key obstacles to scaling diagnostic AI tool deployment for X-ray and CT.

The details:

• 66 hospital trusts analyzed for AI chest tool adoption

• One-third had not implemented 18 months after expected finish

• Barriers: governance, IT, contracts, workload, clinician skepticism

• Solutions: focused training, ongoing management, national leadership

Key takeaway: Understanding operational, technical, and human barriers is essential to move AI from pilot to routine radiology practice.

🧮 Stanford, Mayo: LLMs Flag Critical Findings in Radiology Reports

RadAI Slice: Stanford and Mayo researchers show LLMs can identify urgent findings in radiology reports using few-shot prompting.

The details:

• LLMs achieved up to 90.1% precision, 98.3% recall in detecting critical findings

• Tested on 252 MIMIC-III reports and 180 external chest x-ray reports

• Few-shot prompting reduces annotation burden

• Further EHR integration and refinement needed for deployment

Key takeaway: General-purpose LLMs could automate urgent report triage, but clinical deployment will require robust validation and integration.

Comprehensive Oral Health Hyperspectral Dataset (COH-HSI) (2025-09-11)

Modality: HSI | Focus: Oral cavity, mucosa | Task: Segmentation, classification

• Size: 1,130,751 hyperspectral cubes from 226 patients; 5 annotated images per patient

• Annotations: Pixel-level segmentations; 20+ intraoral classes (mucosa, tooth, palate, etc.), XML and semantic maps

• Institutions: University Medical Center Mainz, University of Applied Sciences Mainz

• Availability:

  Request-only (link)

• Highlight: First large-scale, annotated in vivo endoscopic HSI dataset for oral cavity; optimized for Python workflows

MMOral (2025-09-11)

Modality: X-ray | Focus: dental, jaw | Task: VQA, report generation

• Size: 20,563 panoramic X-rays from ~20,000 patients

• Annotations: 49 categories: bounding boxes, anatomical/structural attribution, 1.3M instruction-following QA pairs, 41k reports, segmentations, dialogues

• Institutions: The University of Hong Kong, Hong Kong University of Science and Technology, et al.

• Availability:

  Public (https://github.com/isbrycee/OralGPT)

• Highlight: First large-scale multimodal dataset for panoramic dental X-ray AI; 1.3M instructions with fine-grained dental labels and evaluation benchmark.

Veriserum (2025-09-05)

Modality: X-ray (dual-plane fluoroscopy) | Focus: Knee, implants | Task: 2D/3D registration, segmentation

• Size: 110,990 images, 10 implant pairings, 1,600 trials

• Annotations: Automated ground-truth pose for all images, 200 images with manual registration

• Institutions: ETH Zürich, Institute for Biomechanics

• Availability:

  Public (link)

• Highlight: First open dual-plane fluoroscopic dataset for knee implants; includes both calibration tools and pose registrations

🏛️ FDA Clearances

• K251533 - Rapid Obstructive Hydrocephalus (Rapid OH) is FDA-cleared (CT) for urgent triage of hydrocephalus, supporting fast brain imaging workflow.

• K251408 - OsteoSight™ Hip (v1) is FDA-cleared to opportunistically assess low bone density from x-ray, aiding osteoporosis risk detection.

• K251656 - Careverse CoronaryDoc is FDA-cleared software for automated coronary image segmentation on CT, supporting cardiac workflow.

• Explore last week's 9 radiology AI FDA approvals.

📄 Fresh Papers

• doi:10.1016/j.radonc.2025.111133 - A multicenter deep learning model (n=485, 4 hospitals) accurately predicts pathologic response and survival in esophageal cancer on CT.

• doi:10.1186/s41747-025-00633-7 - A two-center, externally validated deep learning model segments cartilage tumors of long bones on MRI with mean Dice 0.83.

• doi:10.1007/s00234-025-03745-4 - Meta-analysis of 9,030 patients: AI classification of stroke time using CT radiomics (AUC 0.89) outperformed DWI-FLAIR reading.

• doi:10.1016/j.acra.2025.08.026 - Meta-analysis of 43 studies (n=9,624) finds AI and radiomics models diagnose muscle-invasive bladder cancer with AUC 0.92 and 86% sens/spec.

• Browse 171 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• Bayer will discontinue Calantic and Blackford, removing a major player from the radiology AI platform space and fueling market uncertainty.

• AI-powered MRI reconstruction can speed scans, enhance quality, and reduce radiologist burnout, with up to 12x faster cardiac MRI possible.

• AI using routine mammograms and age matched traditional CVD risk tools, potentially enabling dual-purpose breast screening.

• AI flagged unsuspected lung nodules in chest x-rays, illustrating unforeseen screening potential in routine imaging.

• Researchers propose five foundational features for 'generalist' radiology AI to improve efficiency and reduce fragmentation.

• Read 16 fresh radiology AI news articles for last week.

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.